Related Experiment Video
Updated: Sep 26, 2025
06:25
Author Spotlight: Assessment of Visual Acuity in Central Vision Loss Through Motion-Based Peripheral Vision Testing
Published on: February 23, 2024
737
Video Person Re-Identification with Frame Sampling-Random Erasure and Mutual Information-Temporal Weight Aggregation
1Information and Communication Engineering, Electronics Information Engineering College, Changchun University of Science and Technology, Changchun 130022, China.
Sensors (Basel, Switzerland)
|April 23, 2022
Summary
This study introduces a novel person re-identification (re-ID) method using frame sampling with random erasure and temporal feature aggregation. The approach enhances accuracy in video surveillance despite occlusion and clutter.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Machine Learning
Background:
- Video surveillance systems face challenges with partial occlusion and background clutter, reducing person re-identification (re-ID) accuracy.
- Existing re-ID methods struggle to maintain performance under adverse visual conditions.
Purpose of the Study:
- To develop an improved video-based person re-ID method robust to occlusion and clutter.
- To enhance the accuracy and generalization ability of person re-ID models in complex surveillance scenarios.
Main Methods:
- Implemented a frame sampling-random erasure (FSE) technique for data enhancement to mitigate occlusion effects.
- Utilized a ResNet-50 network for extracting global and partial features, fused into frame-level representations.
- Developed a mutual information-temporal weight aggregation (MI-TWA) module to combine partial and global features across time, learning discriminative sequence features.
Main Results:
- Achieved high performance on benchmark datasets: MARS (mAP 82.4%, Rank-1 86.4%), DukeMTMC-VideoReID (mAP 94.1%, Rank-1 94.8%), and PRID-2011 (mAP 95.3%, Rank-1 95.2%).
- The proposed FSE and MI-TWA methods significantly improved re-ID accuracy compared to baseline approaches.
Conclusions:
- The proposed person re-ID method effectively addresses challenges posed by occlusion and background clutter in video surveillance.
- The combination of data enhancement and advanced feature aggregation techniques leads to more accurate and robust person re-identification.
