Approaches to Improve the Quality of Person Re-Identification for Practical Use

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person re-identification (Re-ID) is crucial for surveillance and security.
• Existing Re-ID algorithms often face practical limitations due to complexity and poor input data.
• There is a need for efficient and robust Re-ID solutions applicable to real-world scenarios.

Purpose of the Study:

• To propose practical approaches for improving person Re-ID quality without increasing computational complexity.
• To introduce a novel Filter Module for pre-filtering input data in Re-ID systems.
• To develop an automated data collection strategy for self-supervised pre-training to enhance neural network generality.

Main Methods:

• Development of a Filter Module to preprocess images before Re-ID analysis.
• Implementation of a fully automated data collection strategy for surveillance camera footage.
• Application of self-supervised pre-training using the collected data for neural network generalization.

Main Results:

• The Filter Module improved baseline Re-ID performance by 2.6% (Rank1) and 3.4% (mAP) on the Market-1501 dataset.
• Self-supervised pre-training enhanced cross-domain upper-body Re-ID by 1.0% (Rank1 and mAP) on the DukeMTMC-reID dataset.
• Proposed methods offer practical improvements with minimal computational overhead.

Conclusions:

• The proposed Filter Module and self-supervised pre-training strategy effectively enhance person Re-ID performance.
• These methods address practical challenges in Re-ID, improving robustness and accuracy.
• The research contributes to more efficient and reliable person re-identification systems for smart city surveillance.