Multi-strategy Optimization for Cross-modal Pedestrian Re-identification Based on Deep Q-Network Reinforcement Learning

Abstract

Cross-modal pedestrian re-identification (C-ReID) is a crucial task in computer vision, aiming to match pedestrian identities across different modalities of data. This paper proposes a reinforcement learningbased framework, RLCMPRF, to tackle the challenges of modality variability, data diversity, annotation difficulties, and optimal strategy selection. RLCMPRF uses deep Q-network (DQN) reinforcement learning to dynamically select the best feature extraction and matching strategies, ensuring robustness against these challenges. We introduce a dual-stream network to process multimodal images, followed by a feature fusion layer for integration. The DQN-based strategy learning is complemented by a reward function designed to optimize matching accuracy, speed, and robustness. Experimental results demonstrate that RLCMPRF outperforms state-of-the-art methods based on deep learning, attention mechanisms, meta-learning, and generative adversarial networks. RLCMPRF achieves a success rate of 82% and an average cumulative reward of 150, showing improvements in convergence speed and generalization ability across multiple datasets.