Related Experiment Video
Updated: Aug 19, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
605
End-to-End Network for Pedestrian Detection, Tracking and Re-Identification in Real-Time Surveillance System.
Mingwei Lei1,2, Yongchao Song1, Jindong Zhao1
1School of Computer and Control Engineering, Yantai University, Yantai 264005, China.
Sensors (Basel, Switzerland)
|November 26, 2022
Summary
This study presents an integrated system for real-time pedestrian detection, tracking, and re-identification in surveillance video. The unified network achieves high accuracy and efficiency for identifying specific individuals across multiple cameras and times.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Machine Learning
Background:
- Surveillance video analysis is crucial for security and investigation but faces challenges in real-time accuracy.
- Existing systems often require separate modules for detection, tracking, and re-identification, limiting efficiency.
- High demands for real-time performance and accuracy necessitate integrated solutions.
Purpose of the Study:
- To develop a unified, end-to-end network for simultaneous pedestrian detection, tracking, and re-identification.
- To enhance real-time capabilities and accuracy in identifying specific pedestrians within surveillance footage.
- To improve the efficiency of surveillance systems by combining multiple functionalities.
Main Methods:
- Integrated YOLOv5 architecture with a new track branch for end-to-end training.
- Employed weighted bi-directional feature pyramid network (BiFPN) for enhanced pedestrian detection.
- Modified Deepsort tracker with Noise Scale Adaptive (NSA) Kalman filter and updated matching strategy for robust tracking.
- Adapted Fastreid network for accelerated pedestrian re-identification.
Main Results:
- Achieved 97% mean Average Precision (mAP) for pedestrian detection.
- Demonstrated high tracking performance with 98.3% MOTA and 99.8% MOTP on the MOT16 dataset.
- Obtained 77.3% mAP for pedestrian re-identification on the VERI-Wild dataset.
Conclusions:
- The proposed unified network effectively balances real-time performance and accuracy for pedestrian analysis.
- The integrated approach offers significant improvements in precise localization and real-time detection of specific pedestrians.
- This framework provides a valuable solution for complex surveillance applications requiring multi-camera and temporal analysis.
