Academic Databases
Explore Our Semantic Databases
STEM & Tech
Math, Physics, CS & Engineering...
Economics
Macro, Micro, Econometrics...
Business
Management, Marketing, Finance...
Health & Life Sciences
Biology, Medicine, Neuroscience...
Humanities & Social Sciences
Psychology, Law, Philosophy, History...
General Studies
General & interdisciplinary studies...
Explore Topics
Features

Real-Time Object Tracking with YOLOv8 and DeepSORT | Vehicle Counting Tutorial

Added: 2026-08-07

Deep Dive

Proficiency in Python programming, particularly with computer vision libraries like OpenCV and numerical processing libraries like NumPy.

OpenCV Python Tutorial For Beginners 1 - Introduction to OpenCV

OpenCV Python Tutorial For Beginners 1 - Introduction to OpenCV

645.6K

OpenCV (Open Source Computer Vision) is an open-source, cross-platform library originally developed by Intel and now maintained by Itseez, designed for real-time computer vision tasks using Python, C, or C++. Computer vision enables machines to see and process visual data like humans, recognizing objects such as faces in images. Digital images are represented as matrices where grayscale images contain one channel (intensity) and colored images contain three channels (Red, Green, Blue). OpenCV integrates with NumPy, a powerful numerical library for array processing, enabling efficient image manipulation and analysis.

Click to watch →

What is OpenCV with Python🐍 | Complete Tutorial [Hindi]🔥

What is OpenCV with Python🐍 | Complete Tutorial [Hindi]🔥

57.9K

OpenCV (Open Source Computer Vision) is a free, open-source library written in C/C++ that enables image and video processing operations including object classification, identification, tracking, feature matching, and motion analysis; it can be installed in Python using pip install opencv-python and accessed through the cv2 module, with essential prerequisites including Python, Anaconda, Jupyter Notebook, NumPy, SciPy, and Matplotlib libraries.

Click to watch →

Don't Buy Scanner! Build your own smart scanner using 25 lines of python || opencv || Python Project

Don't Buy Scanner! Build your own smart scanner using 25 lines of python || opencv || Python Project

94.1K

This segment explains the technical process of converting images to numpy arrays for processing. The instructor demonstrates opening images using the requests library, converting to bytes with image.read(), and creating numpy arrays using np.array() with data type np.uint8. The video then covers displaying images using OpenCV's cv2.imshow() function, with cv2.waitKey() controlling display timing. The segment explains that OpenCV images are stored in BGR format by default, and demonstrates converting to RGB using cv2.cvtColor() with cv2.COLOR_BGR2RGB flag. The instructor also shows how to display grayscale images by passing 0 as the flag to cv2.cvtColor().

Click to watch →

OpenCV Object Detection in Games Python Tutorial #1

OpenCV Object Detection in Games Python Tutorial #1

308.9K

To get started with OpenCV in Python, install it using pip by running 'pip install opencv-python'. After installation, import the library using 'import cv2' (commonly aliased as 'cv'). Numpy is automatically installed alongside OpenCV and is extensively used when working with OpenCV data. A typical Python file setup includes importing both libraries: 'import cv2 as cv' and 'import numpy as np'.

Click to watch →

INTRODUCCIÓN - [Curso Básico de la librería NumPy de Python]

INTRODUCCIÓN - [Curso Básico de la librería NumPy de Python]

38.1K

Numpy is a fundamental Python library for numerical computing that provides efficient array operations, addressing Python's native list limitations by offering memory-contiguous arrays that are up to 50 times more efficient for matrix and vector calculations, with applications spanning mathematics, artificial intelligence, neural networks, physics, and computer vision.

Click to watch →

Deep Dive

Fundamental understanding of Deep Learning and Convolutional Neural Networks (CNNs), specifically how object detection models localize and classify objects.

MIT 6.S191 (2024): Convolutional Neural Networks

MIT 6.S191 (2024): Convolutional Neural Networks

157.8K

Object detection requires localizing and classifying multiple objects with varying positions and sizes. Naive approaches generate thousands of candidate boxes and classify each separately, but this is computationally intractable and loses contextual information. Faster R-CNN solves this by learning region proposals directly using a neural network (region proposal network), followed by shared convolutional processing and individual classification heads. All components train end-to-end with backpropagation, allowing gradients to improve region proposal quality alongside classification accuracy. This represents a paradigm shift from heuristic-based to learned region proposal systems.

Click to watch →

F23 Lecture 12: Deep Neural Networks, Convolutional Networks IV

F23 Lecture 12: Deep Neural Networks, Convolutional Networks IV

509

To locate objects, CNNs predict bounding box coordinates alongside classification. After flattening final layers, two separate outputs predict class existence and box corners. Training uses cross-entropy for classification and L2/intersection-over-union loss for coordinates. The same features serve both tasks. This enables applications like skeleton tracking and pose estimation. CNNs are invariant to position but not orientation or scale—achieving full invariance requires transforming filters for each desired transformation, creating many output channels per filter.

Click to watch →

27. Phát hiện đối tượng - CNN - [Convolutional Neural Networks]

27. Phát hiện đối tượng - CNN - [Convolutional Neural Networks]

174

Object detection combines image classification with localization by training a CNN to output both class labels and bounding box coordinates (bx, by, bh, bw) for detected objects, using a loss function that penalizes errors in both classification and localization parameters.

Click to watch →

CS231n Winter 2016 Lecture 8 Localization and Detection-wFG_JMQ6_Sk.mp4

CS231n Winter 2016 Lecture 8 Localization and Detection-wFG_JMQ6_Sk.mp4

21.5K

Object detection and localization are fundamental computer vision tasks that extend classification by identifying and locating objects within images. Localization involves predicting bounding boxes for a fixed number of objects using regression, while detection handles variable numbers of objects through region proposals. The evolution from AlexNet to ResNet demonstrates how deeper networks improve detection performance. Key approaches include sliding window methods, region proposal networks (RPN), and ROI pooling, culminating in architectures like Faster R-CNN that achieve real-time detection by integrating region proposal generation with convolutional feature extraction.

Click to watch →

Object Detection

Object Detection

126

CNNs for object detection consist of convolutional layers for feature extraction, fully connected layers for classification, and additional layers for localization. The input image is processed through convolutional layers to extract features, then passed to fully connected layers. Two parallel fully connected layers are used: one for class prediction and one for bounding box coordinate prediction. The model outputs class scores (confidence for each class) and bounding box coordinates. A combined loss function includes classification loss and localization loss, trained to minimize the difference between predictions and ground truth.

Click to watch →

Deep Dive

Conceptual knowledge of the YOLO (You Only Look Once) architecture and the mechanics of single-stage object detectors.

YOLO: You only look once (How it works)

YOLO: You only look once (How it works)

50.1K

YOLO (You Only Look Once) is a single-stage object detection model that processes images in one forward pass through a modified GoogleNet architecture, producing a 7x7x30 tensor where each cell contains two bounding boxes with coordinates, dimensions, confidence scores, and 20 class probabilities; the model applies Non-Maximum Suppression to eliminate overlapping detections and selects the highest-scoring bounding box per class for final output, achieving high speed (up to 155 fps) with fewer false positives than R-CNN variants.

Click to watch →

허훈 - YOLO(you only look once) 논문 리뷰

허훈 - YOLO(you only look once) 논문 리뷰

7.1K

YOLO is a real-time object detection model that processes images in a single forward pass by dividing the image into a 7x7 grid, where each grid cell predicts two bounding boxes and their corresponding class probabilities, enabling simultaneous detection and classification while achieving approximately 2x higher accuracy than previous two-stage methods like R-CNN, though it struggles with small objects and multiple objects per cell.

Click to watch →

YOLO V1 - YOU ONLY LOOK ONCE || YOLO OBJECT DETECTION SERIES

YOLO V1 - YOU ONLY LOOK ONCE || YOLO OBJECT DETECTION SERIES

60.4K

Traditional object detection relies on two-stage approaches like R-CNN and Faster R-CNN, which first generate region proposals using RPN and then classify them in a second stage. These methods achieve only ~5 FPS on GPUs, lack end-to-end training convenience, and struggle with domain generalization. YOLO fundamentally reframes detection as a single-stage regression problem, performing box prediction and classification in one forward pass. The input image is resized to 448×448 and divided into a 7×7 grid, with each cell responsible for predicting one object based on whether the object's center falls within that cell. This architectural shift enables dramatically faster inference while maintaining competitive accuracy.

Click to watch →

You Only Look Once: Unified, Real-Time Object Detection

You Only Look Once: Unified, Real-Time Object Detection

133.1K

YOLO is a unified, real-time object detection system that processes images in a single neural network pass by dividing the image into a grid and having each cell predict bounding boxes and class probabilities simultaneously, achieving 45 frames per second while maintaining competitive accuracy of approximately 69 mAP on Pascal VOC, significantly faster than previous methods like R-CNN and Fast R-CNN.

Click to watch →

L49: YOLO: You only look once | object detection explained | fast & accurate models

L49: YOLO: You only look once | object detection explained | fast & accurate models

676

YOLO (You Only Look Once) is a single-stage object detection algorithm that divides an image into an S×S grid, where each cell predicts B bounding boxes with coordinates (XY, width, height), confidence scores indicating object presence, and class probabilities; the network outputs all predictions in a single forward pass, making it significantly faster than two-stage detectors like R-CNN, with a loss function that uses different weights (Lambda) for coordinate predictions, confidence scores, and class probabilities, and applies square root transformation to penalize small bounding box errors more heavily.

Click to watch →

Deep Dive

Basic familiarity with state estimation and data association algorithms, such as Kalman Filters and the Hungarian method, which underpin tracking systems.

Multi Object Tracking Tutorial: part 1 by Student Dave

Multi Object Tracking Tutorial: part 1 by Student Dave

33.8K

Multi-object tracking involves maintaining the identity of multiple moving objects over time by combining detection (identifying object positions), prediction (using Kalman filters to forecast future positions), and assignment (using the Hungarian algorithm to optimally match predicted positions with detected positions, enabling robust tracking even when objects cross paths, occlude each other, or temporarily disappear).

Click to watch →

Wolfgang Koch - Lecture 2, Stefano Coraluppi - Lecture 1

Wolfgang Koch - Lecture 2, Stefano Coraluppi - Lecture 1

328

Filtering estimates state from measurements, with optimal estimators depending on cost functions: conditional mean for squared error, conditional median for absolute error, conditional mode for maximum likelihood. The Kalman filter provides optimal estimates for linear systems with Gaussian noise, written recursively for computational efficiency. Multi-target filtering assumes knowledge of object count and perfect detection, with challenges in determining which measurement belongs to which target. Statistical nearest neighbor assigns measurements sequentially based on Mahalanobis distance. Global nearest neighbor minimizes overall cost across all tracks. Probabilistic Data Association (PDA) assigns weighted combinations of measurements to tracks. Multi-Hypothesis Tracking (MHT) uses hard association but waits for additional scans, considering multiple hypotheses about previous scans.

Click to watch →

Слежение за объектами: алгоритмы привязки к целям и построения трасс

Слежение за объектами: алгоритмы привязки к целям и построения трасс

1.3K

Object tracking estimates object positions in modern systems like surveillance and autonomous vehicles without human intervention. The Kalman filter is the fundamental mathematical tool, using motion models and sensor measurements to estimate states. Sensors introduce measurement noise, while motion models may not perfectly match real movement. The filter balances prediction uncertainty against measurement noise, finding optimal estimates. Gaussian distributions model uncertainties, and the filter combines predictions and measurements to produce corrected estimates with lower variance. Motion models (constant velocity, acceleration, turn rate) must be selected based on target behavior, but single models fail for maneuvering targets. Interactive Multiple Models (IMM) addresses this by running multiple Kalman filters simultaneously, each with its own motion model. Data association determines which detection belongs to which track, with simple nearest neighbor algorithms failing for closely spaced targets. Joint Probabilistic Data Association (JPDA) assigns weights based on likelihood. Track management handles dynamic scenarios where objects appear/disappear and sensors produce false detections.

Click to watch →

Multi Object Tracking Tutorial: part 2 by Student Dave

Multi Object Tracking Tutorial: part 2 by Student Dave

21.7K

This tutorial demonstrates a multi-object tracking approach that combines a Kalman filter for predicting object positions with the Hungarian algorithm for matching predictions to detections. The process involves first detecting objects using a Gaussian filter followed by a Mexican hat filter to identify local maxima as centers of mass, then using the Kalman filter to predict their next positions, and finally applying the Hungarian algorithm to optimally assign predicted positions to actual detections in subsequent frames, enabling robust tracking even when objects overlap or occlude each other.

Click to watch →

SenSys 2015 - Accurate Positioning via Cross-Modality Training

SenSys 2015 - Accurate Positioning via Cross-Modality Training

144

The data association problem arises when observations are anonymous (no target IDs). The solution involves modeling associations as variables in the system. Rao-Blackwellized particle filtering decomposes this problem: first estimate the association posterior distribution using particle filtering, then estimate target states conditioned on associations using analytical Kalman filtering. This decomposition reduces computational complexity in high-dimensional state spaces. Each particle maintains a bank of Kalman filters tracking all targets, with particles representing realizations of the system that track association history. The approach samples association distributions and uses Kalman filters for state estimation, enabling robust tracking despite occlusions and crossing paths.

Click to watch →

Setup & GitHub

Environment Setup

Directory Navigation

Download Weights

Initial Test Run

DQ & Colors

Bounding Boxes UI

Trails & Tracking

Detection & Classes

Counting Logic

Setup & GitHub

0:12

Playing Section

1
Open PyCharm and create a new project directory.
2
Clone the YOLOv8 Deep Sort repository from GitHub.
3
Set up the environment and install all required dependencies.
4
Note that CPU and GPU implementations require identical steps.

Deep Dive

Exploration of state-of-the-art tracking algorithms like ByteTrack or BoT-SORT to compare performance, accuracy, and handling of occlusions.

Object Tracking YOLOv8 and ByteTrack (Player Tracking and ByteTrack Algorithm Explained)

Object Tracking YOLOv8 and ByteTrack (Player Tracking and ByteTrack Algorithm Explained)

12.9K

ByteTrack is a multi-object tracking algorithm that improves upon traditional trackers by utilizing both high and low detection thresholds from an object detector (like YOLOv8), combining Kalman filter predictions with the Hungarian algorithm for association, and managing tracks through parameters including high/low thresholds, match threshold, track buffer, and new track threshold to achieve superior MOTA performance while handling occlusions effectively.

Click to watch →

Realtime Object Tracking | Computer Vision Project | Euron

Realtime Object Tracking | Computer Vision Project | Euron

496

Object tracking in computer vision involves detecting objects using bounding boxes and then maintaining their identification across consecutive video frames using tracking algorithms. The process requires first applying an object detection model (such as YOLO 11) to identify objects in each frame, then passing these detection results to a tracking algorithm that assigns unique IDs to each detected object and follows them throughout the video. Two commonly used tracking algorithms are ByteTrack and BoT-SORT, where ByteTrack offers faster performance but cannot handle occluded objects (losing track when objects disappear), while BoT-SORT provides better accuracy with occluded objects at the cost of slightly slower processing. The choice between these algorithms depends on the specific application requirements—applications with predictable object movement (like manufacturing systems) may prioritize speed, while applications with unpredictable occlusions (like traffic monitoring) require the more robust tracking capabilities of BoT-SORT.

Click to watch →

How to Track Objects with YOLO and ByteTrack/BoT-SORT Trackers

How to Track Objects with YOLO and ByteTrack/BoT-SORT Trackers

282

This video demonstrates how to combine YOLO object detection models with ByteTrack and BoT-SORT tracking algorithms for real-time object tracking in videos. The process involves installing necessary libraries (ultralytics, opencv, pytorch), loading a pre-trained YOLOv8 nano model, reading video input, applying the chosen tracker algorithm, extracting detection information including bounding box coordinates, confidence values, class labels, and track IDs from the results, and displaying the output with FPS values. The video shows that ByteTrack achieves higher FPS (around 30) compared to BoT-SORT (around 22-23), and emphasizes that GPU support significantly improves performance (20-30x faster than CPU). Different YOLO model sizes (nano, medium, small) offer trade-offs between speed and accuracy.

Click to watch →

Python: Real-time Multiple Object Tracking (MOT) with Yolov3, Tensorflow and Deep SORT [FULL COURSE]

Python: Real-time Multiple Object Tracking (MOT) with Yolov3, Tensorflow and Deep SORT [FULL COURSE]

117.8K

The SORT algorithm uses Kalman filters for motion prediction and Hungarian methods for data association. It maintains object models with spatial information and motion predictions, associating detections via IoU distance. However, SORT struggles with occlusions due to its simple motion model. DeepSORT improves this by incorporating appearance features extracted via CNNs, enabling better handling of occlusions and achieving higher accuracy. The three-step implementation includes: object detection, motion prediction/feature generation with Kalman filtering, and tracking with cosine distance, IoU distance, and state distance calculations.

Click to watch →

YOLO11 Object Tracking | People Tracking Using Byte Track and Bot SORT

YOLO11 Object Tracking | People Tracking Using Byte Track and Bot SORT

2.3K

Object tracking involves detecting objects in video frames and following them across consecutive frames using unique IDs, with modern algorithms like ByteTrack (fast but less accurate for occlusions) and BoT-SORT (slower but better for occlusion and re-identification scenarios) addressing different application requirements.

Click to watch →

Deep Dive

Model optimization and deployment techniques using TensorRT, ONNX, or OpenVINO to achieve high-frame-rate inference on edge devices.

Anomalib: A Deep Learning Library for Anomaly Detection - #OpenCV Live Ep 112

Anomalib: A Deep Learning Library for Anomaly Detection - #OpenCV Live Ep 112

12.2K

OpenVINO provides multiple optimization techniques for improving model performance: Model Optimizer for general model transformations, Neural Network Compression Framework (NNCF) for post-training quantization and quantization-aware training. These techniques convert full-precision models to lower precision formats (FP32 to INT8) while maintaining acceptable accuracy, resulting in dramatically reduced inference latency. The optimized models can then be deployed using OpenVINO Runtime, achieving inference speeds ranging from several frames per second on edge devices to over 3,000 frames per second on high-performance Xeon processors, depending on the hardware platform.

Click to watch →

How to Choose the Best Ultralytics YOLO11 Deployment Format for Your Project | TensorRT | OpenVINO 🚀

How to Choose the Best Ultralytics YOLO11 Deployment Format for Your Project | TensorRT | OpenVINO 🚀

1.1K

After training models, multiple optimization frameworks exist for production deployment. These frameworks increase inference speed by 2-5x while maintaining accuracy. PyTorch serves as the base framework, with TorchScript offering better production support and C++ compatibility. ONNX acts as a standard format for converting models between frameworks. TensorRT optimizes for Nvidia hardware, achieving 3-4x speedup with FP16/INT8 precision. OpenVINO targets Intel hardware for CPU inference. CoreML serves Apple devices with M1 chip optimization. For mobile and edge devices, TensorFlow Lite and NCNN provide ARM system optimization. Framework selection depends on target hardware, deployment requirements, and performance needs.

Click to watch →

Intro to Object Detection & Computer Vision Pipelines - AI Workshop

Intro to Object Detection & Computer Vision Pipelines - AI Workshop

174

Model optimization techniques include: quantization (reducing precision from float32 to int8), pruning (removing less important neurons), and using specialized runtimes like OpenVINO (Intel), TensorRT (NVIDIA), or ONNX Runtime. These techniques involve trade-offs between accuracy, speed, and resource usage. For edge deployment, lightweight models like YOLO are preferred, and quantization can achieve 60% compression with minimal accuracy loss.

Click to watch →

NVAITC Webinar: Deploying Models with TensorRT

NVAITC Webinar: Deploying Models with TensorRT

20.8K

TensorRT is NVIDIA's software development kit for high-performance deep learning inference that addresses two key challenges: high performance and platform portability across edge devices, autonomous driving platforms, and data center GPUs. It implements five core optimization technologies: kernel fusion (combining multiple kernels into single executions to reduce overhead), precision calibration (using lower precision during inference to reduce model size and latency), kernel auto-tuning (selecting optimal implementations based on configuration), dynamic tensor memory (allocating memory only for tensor duration), and multi-stream execution (enabling parallel inference for multiple clients). The TRTorch library provides a straightforward conversion pipeline that abstracts these optimizations, allowing developers to convert trained PyTorch models into optimized TensorRT modules with minimal code, achieving significant speedups (1.5-6.4x) while maintaining model accuracy.

Click to watch →

Intro to On-device Machine Learning (TF Fall 2020 Updates)

Intro to On-device Machine Learning (TF Fall 2020 Updates)

23.6K

Edge devices have limited CPU, memory, and battery power, making model optimization critical. Three main optimization methods exist: (1) Mobile-optimized architectures like MobileNet V3, which is 40x faster and 8x smaller than Inception while maintaining reasonable accuracy; (2) Quantization, converting 32-bit float weights to 8-bit integers, reducing model size by 4x with minimal accuracy loss (~1%); (3) Pruning, setting weights to zero to shrink models, with 50% pruning causing only ~1% accuracy drop. These methods are compatible and can be combined for maximum efficiency. Hardware accelerators like CPUs and Edge DPUs further speed up inference. The key trade-off is between accuracy, model size, and inference speed.

Click to watch →

Deep Dive

Implementation of advanced spatial-temporal analytics, such as vehicle speed estimation, trajectory prediction, and virtual tripwire analysis.

How to Configure Tripwire Detection | Web Browser & NVR Tutorials - Diamond Series

How to Configure Tripwire Detection | Web Browser & NVR Tutorials - Diamond Series

9.5K

This section demonstrates the complete tripwire analytics workflow including real-time triggering and event retrieval. The tutorial shows how to configure tripwire direction (left-to-right, right-to-left, or both), select target filters (human or motor vehicle), and draw the analytics line. It then demonstrates the efficient AI Search method for retrieving triggered events: navigating to AI > AI Search > IVS, selecting the camera, setting time range, choosing Tripwire event type, and filtering by target. The tutorial also covers the less efficient timeline search method for comparison. The demonstration includes live view playback showing the analytics line flashing red when triggered, with the system displaying event type and target information.

Click to watch →

Real-time estimation of distributed parameters systems: application to traffic monitoring

Real-time estimation of distributed parameters systems: application to traffic monitoring

334

Virtual trip lines are virtual sensors that tell cell phones where to send measurements. Phones download map tiles and virtual trip line coordinates, then check if their GPS trajectory crosses any line before sending measurements. This framework preserves privacy by preventing identification of individual users through probabilistic measurement sending and avoiding residential areas. Measurements are triggered on major freeways where anonymity is maintained.

Click to watch →

MobilityDB: Hands on Tutorial on Managing and Visualizing Geospatial Trajectories in SQL

MobilityDB: Hands on Tutorial on Managing and Visualizing Geospatial Trajectories in SQL

1.1K

MobilityDB provides sophisticated analysis functions for trajectory data. The speed() function computes varying speed values at each timestamp with linear interpolation between observations. The time_weighted_avg() function calculates average speed by computing the area under the curve divided by duration. Temporal aggregation functions like temporal_count() count vehicles at each instant during specified periods. The cumulative_length() function creates temporal floats showing distance driven so far at every instant, with the final value matching the overall trajectory length for validation purposes.

Click to watch →

Enabling Mobility Data in Multiple Computing Environments

Enabling Mobility Data in Multiple Computing Environments

400

Mobility DB provides native support for spatial-temporal analytics that automatically handles trajectory semantics without requiring manual interpolation. Examples include determining when a bus was within 30 meters of a billboard (spatial buffer intersection) and detecting near-collisions between moving ships (distance predicate). These predicates work by recognizing connected moving points and applying appropriate geometric operations. The system automatically filters out irrelevant data (like trajectories never approaching a target) and returns precise time intervals when conditions are met. This abstraction simplifies complex queries that would otherwise require writing custom interpolation logic in traditional GIS systems.

Click to watch →

Accelerate Your Mastery of Mobility Data Analytics

Accelerate Your Mastery of Mobility Data Analytics

103

Modern mobility analytics combines floating vehicle data (IoT sensor data from connected vehicles providing position, speed, and trip information) with GTFS data (real-time transit vehicle locations and schedules) to enable sophisticated spatial-temporal analysis that reveals traffic patterns, public transit delays, and urban mobility challenges; platforms like xyzt.ai address the challenge of analyzing big spatial-temporal data by integrating BI, GIS, and time series analysis capabilities into self-service tools that allow analysts to extract insights from billions of data points without requiring extensive programming or GIS expertise, transforming the traditional data science workflow where 90% of time was spent on development into a model where analysts can focus 90% of their effort on solving actual business problems.

Click to watch →

Deep Dive

Developing Multi-Camera Multi-Target (MCMT) tracking systems to re-identify and track vehicles across non-overlapping camera views.

Keynote_(Person Re-Identification and Tracking in Multiple Non-Overlapping Cameras)

Keynote_(Person Re-Identification and Tracking in Multiple Non-Overlapping Cameras)

7.7K

Multiple camera tracking addresses surveillance scenarios with varying configurations: fixed overlapping fields requiring handover, fixed non-overlapping fields requiring re-identification, and moving cameras adding viewpoint complexity. Classical approaches solve single-camera tracking first, then address cross-camera associations. The unified framework simultaneously solves within-camera and across-camera tracking through three hierarchical layers: raw detections → short 15-frame tracklets using overlapping constraints → full tracks within cameras via constraint dominant set clustering → cross-camera associations connecting tracks from different cameras with logical constraints (each track belongs to at most one set per camera). This hierarchical approach successfully stitches broken tracks across non-overlapping camera fields of view, demonstrating effectiveness on DukeMTMC dataset with eight fixed cameras and over 2 million frames, achieving significant improvements in tracking metrics for both easy and hard cases.

Click to watch →

Multi-target Tracking in Multiple Non-overlapping Cameras Using Fast-Constrained Dominant Sets

Multi-target Tracking in Multiple Non-overlapping Cameras Using Fast-Constrained Dominant Sets

8.9K

This video presents a novel multi-target tracking framework for multiple non-overlapping cameras that formulates the tracking problem as finding dominant sets in graphs. The approach uses Fast-Constrained Dominant Sets (FCDS), which transforms the combinatorial clustering problem into a linearly constrained quadratic optimization problem, enabling efficient merging of within-camera tracks and across-camera identification. The method achieves approximately 10% improvement in MOTA scores on the DukeMTMC dataset compared to state-of-the-art approaches, while maintaining persistent ID assignment across camera views.

Click to watch →

CVPR 2019 Oral Session 3-1B: Learning, Physics, Theory, & Datasets

CVPR 2019 Oral Session 3-1B: Learning, Physics, Theory, & Datasets

3K

CityFlow is the first city-scale benchmark for vehicle-based multi-target multi-camera tracking and re-identification, containing five scenarios across 40 cameras at 10 traffic intersections, 3.25 hours of video, and 666 unique vehicle identities. Vehicle tracking is more challenging than person-based tracking due to higher intraclass variability from viewing angle changes and higher interclass similarity between car manufacturers. Metric learning with triplet loss combined with cross-entropy achieves best performance. HD maps enable lane-based trajectory prediction outperforming sophisticated LSTM models with social context but no map access.

Click to watch →

The 4th AI City Challenge

The 4th AI City Challenge

1.4K

City Flow is a city-scale benchmark for multi-target multi-camera vehicle tracking and re-identification, addressing the challenge that vehicle-based problems are more difficult than person-based ones due to higher intraclass variability (vehicles appear differently from various angles) and higher interclass similarity (different car models look alike). The benchmark includes five scenarios across 40 cameras at 10 traffic intersections, totaling 3 hours and 15 minutes of video with 660 vehicle identities. Evaluation shows that combining cross-entropy loss with arcface loss achieves the best performance for vehicle re-identification, while spatial-temporal association schemes using Gaussian distributions enable reliable trajectory association across cameras.

Click to watch →

Multi-Target Multi-Camera Tracking of Vehicles by Graph Auto-Encoder and Self-Supervised Camera Lin

Multi-Target Multi-Camera Tracking of Vehicles by Graph Auto-Encoder and Self-Supervised Camera Lin

783

This paper proposes a novel framework for multi-target multi-camera vehicle tracking that addresses challenges like similar vehicle appearances, feature variations, and occlusions by combining a graph auto-encoder for robust feature learning with a self-supervised camera link model that automatically establishes topological and temporal relationships between cameras without manual labeling, achieving state-of-the-art performance on the CityFlow dataset.

Click to watch →

Setup & GitHub

Environment Setup

Directory Navigation

Download Weights

Initial Test Run

DQ & Colors

Bounding Boxes UI

Trails & Tracking

Detection & Classes

Counting Logic

Setup & GitHub

0:12

Playing Section

1
Open PyCharm and create a new project directory.
2
Clone the YOLOv8 Deep Sort repository from GitHub.
3
Set up the environment and install all required dependencies.
4
Note that CPU and GPU implementations require identical steps.

64.1K views85029:09@muhammadmoinfaisalOriginal Release: 2023-01-12

This video demonstrates how to implement real-time object tracking using YOLOv8 for object detection combined with DeepSORT for multi-object tracking, including vehicle counting by detecting when tracking trails intersect predefined lines to count vehicles entering and leaving a highway road.

[00:00:12]hi everyone in this video we will see how we can Implement object tracking using deep sort in Yolo V8 we will be tracking ID as well as Trails we will create an application in this video as well in which we will implement the number of vehicles entering and leaving a highway road so as you have seen the demo at the start of the video so let's move forward in the first step you can see as this is my GitHub repo uh which I have created for YOLO V8 deep sort tracking so first of all I will go on this GitHub repo into my IDE I will be using pycharm for the implementation of this project so as I don't have a GPU so you can also implement it on CPU I will be showing demo on CPU so if you have a GPU you will follow the same steps so no step will be changed you will follow each of the same step if you have a GPU based system if you don't have a GPU based system then I also don't have a GPU based system so I will be following the same steps so here is the collab filing as well you can Implement in Google collab as well it's a single click solution so I will be showing you the implementation of pycharm ID in this video I will be explaining the complete code as well so please watch the complete video and so I hopefully at the end of the video uh your very much Concepts will be cleared so let's start the implementation so first of all I will create up icon I will open python so this is uh my GitHub repo I have just published 10 hours back so let's look into more details so just open the pycharm okay in the meanwhile as the python is loading let's create a folder here so we can name it uh leaps or nothing in tutorial okay that's nice something okay so first of all I will go to files this is previously opening project I was just checking few things go to new section okay no sorry I'll go to open and then what I have created yeah this is the file this is in this file trust project click on trust project this window yes so let's move forward so first of all uh not nothing to worry about just go to the terminal just open the terminal I am using pycharm us you can use BS code or any other ID which you like to work on so no worries if you are working on some any other ID so first of all uh we need to upload the GitHub repository so here all the steps are mentioned as well so let's just follow the mention uh the steps I have mentioned in the GitHub repository so first of all the step step is to clone the repository I will be explaining your code as well don't worry so just go here and just Ctrl D and press enter nothing to worry about so it might takes two to three minutes so let's wait okay so now we have drawn the GitHub repository okay so in the next step uh we need to go to our clone folder so just copy this and just go to the font folder so now we are in the control called uh clone folder now we need to install the requirements as well as install Ultra latest so install requirements right to install minus and dot txt let's install all the requirements uh here we have this requirements file you can see over here now install Alternatives as well like you can see that the next step after requirements is to install electrolytics installed and from Analytics so it might take a few process so as I have already done many work working on your search appearing like requirements already satisfied like uh these species are all already installed in my Python 3 Library so it's not installed on your site it might take few minutes so don't worry about it so now setting the directory so okay so now we are setting it directly as yellow V8 detect so let me show you what it is so this is Yolo and let us close this first and this is the A8 folder and then we go to the detect folder so first of all let's look at a tracking strip so this is the tracking dot by file which I am redirecting my directory to so for example currently my directory is set as configure sentence just setting this [Music] um give me a second so just click on ok now I'll just continue so this is a tracking dot Pi file over here I am just redirecting uh to this folder like the tag folder which contains our tracking dot file so currently I am over here your low vid sorry in the main folder so to redirect over here I will go to CD YOLO Dash to Turk Dash V8 sorry don't this is the V8 and detect is the subfolder of the video detect and just click on enter okay so now we are here in the detect folder and this is the tracking dot Pi file okay so now we have redirected to this folder so the next step is to download these deep sort files from this Google Drive and place into the uh you know V8 folder so basically the Deep sort uh sort file size are very large so we we cannot uh upload on GitHub of a file above like 25 MB so these are very these files are flies have above 100 MB so we would prefer I so therefore I have uploaded these files on the Google right so you can download these files from Google Drive from here and just place them into the folder like that I will mention so it might be a few minutes let us go let us okay that's weird so okay so before you download it okay okay it might take a few minutes so now just go to this folder uh this folder is present over okay and just click on extract files okay this is the folder these are the five so I will just copy it and just go to YOLO is see this is the folder and just I will paste over okay this is my folder and I will just paste these files over here okay so please follow the step which I have explained here don't skip and Skip okay so if I see now these are deep sort by Dodge files appearing over here why we are downloading sort by torch because we are implementing our tracking using deep sort tracking ID as well as streams using deep sort so please remember this point Okay so let's go over here and just uh for to for testing I'm will download the sample video from this driving so directly download the sample video from here okay and now I've downloaded the sample videos here so now just to run this script so if as I'm testing on CPU so it might take time to run the script it will take time to run the script so in the meanwhile the script will be running I will explain you the complete script so what's going on at the back end okay so it's working so in the meanwhile as you can see the under script is being running so let's go to the code so as you can see the script is being running so it might take some time like you can see that um it's working so let's create the till the implementation gets complete and then we will see the output video so in the meanwhile let me explain you the complete code so first of all uh this is the tracking dot Pi script for example this from here uh basically we have imported the best predictor from YOLO uh ultralix's Yellow Engine as well as I'm just importing the annotator colors and save one box although we are you're not using the same unbox over here so but I'm important from the planting Library here I have just imported cue2 so in computer vision library from here I have from this is from these are the libraries which I have imported from the deep sort files these two libraries and from collections I have imported DQ so what we are doing is in DQ is basically DQ is a so DQ is basically uh you can say that let me write out here DQ is basically uh double ended queue oh my God so why we use DQ basically uh we prefer DQ over list you over what it is when we need to perform insertion or pop-up patients defense at the same time so DQ is basically a double ended queue we prefer DQ over list when we need to perform insertion or pop operation over same time so here we are basically where we are using here we uh why we are using DQ is basically uh when we find the bottom center of the boarding box we basically uh after finding the bottom center of the bonding box like for example here we have a card or a drug after funding uh the Border symptom water set bottom center of the bonding box uh we append basically them to uh in the form of DQ like for example the bottom center uh value is 2.8 I will append this value in the DQ so basically uh we can Define the maximum length of DQ as 64 and 32 so this value will be appended into the DQ for example of now the car exists from the frame Grand frame so we need to remove this value from the EQ so okay so basically uh when the cars come into the frame the value of this bottom center edge of the botting bounding box uh is appended into the DQ as the car disappears from the frame we need to remove the value from the DQ of the center Center value from the bottom center of the bondix was value from the DQ so here we are performing insertion and pop-up operation at the same time so therefore we prefer EQ over a list thus we can Define have defined the maximum length of dqs64 so it will not accept values Above 64. So as the values get 64 it will not accept the further values okay so this is here we are using the DQ uh as well so further details will be given so here we are defining the color palette here we are defining DQ and here we are just initializing deep sort of none and here uh this is basically the code here we have initialize all the Deep sort because my tech tracker is deep sort uh in some cases you can use sort alt give them as your tracker buy track as your tracker as well but here I'm preferring a deep sort as my tracker okay so this is all the code where we basically calculate the relative bonding losses for absolute pixel values and this the same quality code is fair as well so this is basically how we compute color for labels for example if the label is scarred so this specific color will appear for the car so basically it's a purple color and if we have a bus it's around pink light pink color and we have different color for the water uh motorbike so if the label is not from these so it will randomly slide color from the palette so here we have defining now basically here what we are doing is we are creating a draw using draw Dash border uh basically we are defining a good UI so let me tell you uh a bit more in it so you can see that this is the car this is the UI like you can see that this is appearing in the rectangular form like you can see that this is in a rectangular form not in the Square so if I want to create this into the square form I can simply use cb2.

[00:14:26]rectangular rectangle but you can see that the UI is in the form of like stylish form like because it's not in the rounded rectangle form like you can see that this is in the form of laundry rectangles so I am just creating this uh UI through uh basically I'm getting a separate function uh to make my UI border so uh better so draw that border and the function which contain which creates a rounded rectangle UI like using this function we create this rounded rectangle UI I'm just talking this there we have written 17 card so you can see that this appears in the rounded form rounded rectangle from so using draw Dash border function we create a rounded rectangle UI okay so now further we have the UI Dash box uh in UI Dash box we just uh write text or what like you can see that okay so this is a 17 or card so basically in UI Dash box we basically pass this label value into the rounded rectangle we have created in the draw Dash border function so just this thing like you can see that we have calling the draw Dash border function from above over here and just putting the text over here into it so okay so here we have the function raw Dash boxes basically here you can see that we are finding the center of the body bottom Edge and then we are just calculating the identities for each of the word detected object and you can see that we have defined the maximum length of DQ uh our double ended Q as 64 it will not accept values above 64 and here we are uh finding the object name what object name is like it's car or truck and here we are just finding the format of our label like we have a unique ID for each of the detected object like this is uh here we have let me show you like you can say that 29 39 30 each uh object or detective object has a unique ID like 29 39 36 38 30. okay so we are assigning a unique ID to each other clean project and in the label we have the object name well like if it is car it is copyright into it is motorbike it is appearing motorbike and then after finding this finding the bottom center of the uh bonding box we are appending this into our double ended queue okay so basically why we are appending to double ended queue because we need to generate Trails because using this we generate Trails let me show you what are trails this is seriously okay so now you can see that uh this line just give me a minute so now you can see that this line over here this line this is our Trace so basically uh our each of the center of the bonding box values will be appended into the data DQ uh double ended queue list and using this uh the CD generator queue like you can see that this is the line we have generated so all because we have appended all the values in the form of data DQ and using Delta DQ we generate this line like you can see over here from the bottom center of the boarding box you can see that this is the bottom center of the body box you can do at the center of the bonding box or the top center of the bounding box I am just focusing on the bottom center of the body box as well okay so here we have done and here we are just creating like you can see the trails like I have shown you the trails over here which I have shown you like these are the trails like you can see that these are the trails like this and this so basically we are creating using CB2 dot line and you can see that we are taking the last value and the current value to generate the trails you can see over here find up data DQ then we have a data predicted deduction prediction function first people do the pre-processing then we apply non-mex regression uh to resize our coming image from the frame by frame and basically we have the video in the form of frame by frame to resize that frame and then further we are basically doing a annotations and here basically we are performing detections one by one like you can see that here I'm just calling the Deep sort function which have I initialized over here I'm just calling this function over there like you can see that this is a deep sort function which I have initialized so I'm just calling this function over here we saw dot update in the form of here and basically now I'm just passing these four values as you can see over here I'm just passing these four values over here uh this is the object ID we are getting as the output here here we have the object ID as well here we have the identities like uh object ID basically help us to find the class name object ID basically has bus 2.4 example we know that at 0 we have person at one we have got so basically if the object ID is zero we know that this is a person object ID is one we know that it is a car or and identities have assign help us to assign unique ID to each of the detected object okay and set dot modern dot names basically is help us to get uh the data the information from the Coco data set and this is the bounded boxes like for example um from where from the points of the bonding boxes like that like you can see that [Music] um okay basically uh this may tells me x one y one x two y two so this tell me these two points of the bonding boxes okay so this this gives me the information about these two points of the washing boxes and this is my current frame which where I am pending all this information so this is all the code here I'm just running the script and So currently we have done this script till this so just stop the script and check what out we have just stop the script over here okay so let me go to here we have the information [Music] so this is the current output we have like you can see that I am not using I am using a nano model so not giving something very perfect prediction so like you can see that these are the trails this is the unique ID and I did achieved up there okay so now moving further as well uh now if you want to uh implement the number of vehicles entering and we are leaving like you can just seen at the start of the demo at the start of the video so you can run this script as similarly you can tell me what topics you like you are facing an issue in such topic I will create a demo videos for these topics as well so we are basically creating a complete computer vision series for Europa and I will be creating two to three lectures every week so it's going to be a very interesting learning so do subscribe the channel so that I will share more content now let us see what happens so as the basically uh let me I will explain this file as well so let us just see so when this Trail intersect this line let me check let us take it plywood change when these two will intersect this line okay now yes okay when this like you can see that when this player intersect with slime it detects like this is the drop one always misdirecting what it is like the object has passed so when this will pass over here it will do the same so let us wait for this as well to so just let us wait for a few minutes it's awesome I'm just waiting to just to pass over here foreign [Music] so now you can see that let's just wait for a few more minutes for this pass as well and as it just intersects okay the count is increased so you can just check this out as well further now let me explain this code to you basically uh in that demo video you have seen the line at the start like when the trails of the car intersect with the line uh basically it blinks and the count increases so here we have defined this line like let me show you for the so that we don't have any doubts let me just give you a minute so this is the line um can I edit it okay so this is the line uh this is a complete line you can see that this is the complete fine Green Line we are talking so when this trace this intersect with this line the count increases so here I have defined this line as here and this is the initialized studied sort and this is all the same board we have discussed the cardboard border let me show you the difference Here Comes This is the same code and this are the intersect CCW and get direction for example uh if that direction is we have to find the directions for Atomic direction is sort or not so for example at least and West we are just focusing on these two direction in this video like South South and North Okay so find this function over here so uh basically if the south is interaction it will increase the count plus one if not in the direction it will also increase the count plus one so for example if that Trails uh if the if the trains are moving in the direction of South it will increase the count as well and if the trails are moving in the direction of North it will increase the count as well let me show you so like you can see that uh industry forgot now South East was this is not his problem okay so for example as you can see that this here we have this South and here we have the North so if the trails move in this direction the count will increase the number of vehicles in leap entering and if the trails come in this direction like you can say this is in the South Direction it will represent the number of vehicles leaving okay so in this way we are increasing the count and here I've just set the UI like here just the uh what is over here is basically I have set all the UI all this UI like you can see that this UI and this UI is being set through this like this UI so I understand the UI and does the same either under detected function as well and nothing further so I hope you have liked this video so do watch this complete video and if you have any issues do let me know in the comment section I will show the answer

Up Next

AI Quantization Explained: Optimize Local LLM Performance

@technovangelist

403.3K views•2024-12-28

Building Real-Time ML Pipelines with Feature Stores and MLOps Frameworks

@ODSCAI

5.1K views•2022-02-20

Bypassing Tor Censorship: Bridges and Pluggable Transport Guide

@Coding_ForEveryone

397 views•2024-06-11

Neural Networks Explained: Math, Layers, and Learning Fundamentals

@3blue1brown

21.9M views•2017-10-05

Related Study Plans & Knowledge Roadmaps

Structured learning paths in Artificial Intelligence

View All Artificial Intelligence Topics

artificial-intelligence

Building Recommenders: Two-Tower & Vector DBs

01Foundations of Python & Machine Learning Basics

02Introduction to Recommender Systems

03Deep Learning & Representation Learning

+ 3 more learning modules

Explore Roadmap

artificial-intelligence

Latent Diffusion: PyTorch, UNet & CLIP

01Deep Learning & PyTorch Foundations

02Convolutional Networks & Autoencoders

03The Math and Mechanics of Diffusion Models

+ 3 more learning modules

Explore Roadmap

artificial-intelligence

NeRFs: 3D Reconstruction & Instant-NGP

01Math & Python Foundations for 3D Vision

02Photogrammetry & Camera Pose Estimation

03Neural Radiance Fields (NeRF) Core Concepts

+ 2 more learning modules

Explore Roadmap

artificial-intelligence

TFT Forecasting: Energy Demand & Deep Learning

01Programming & Data Science Foundations

02Introduction to Machine Learning & Time Series

03Deep Learning & Sequential Models

+ 3 more learning modules

Explore Roadmap