Enriched and Discriminative Human Features for Person Re-Identification Based on Explainable Behaviors of Convolutional Neural Networks

Abstract

Understanding pedestrian behaviors such as their movement patterns in urban areas could contribute to the design of pedestrian-friendly facilities. With the commonly deployed surveillance cameras, pedestrian movement in a wide region could be identified by the person re-identification (ReID) technique across multiple cameras. Convolutional neural networks (CNNs) have been widely studied to automate the ReID task. CNN models equipped with deep learning techniques could extract discriminative human features from images and show promising ReID performance. However, some common challenges such as occlusion and appearance variation are still unsolved. Specifically, our study infers that over-relying on discriminative features only may compromise ReID performance. Therefore, this paper proposes a new model that extracts enriched features, which is more reliable against those ReID challenges. By adding a feature dropping strategy during model training, our model learns to focus on rich human features from different body parts. Moreover, this paper presents an explainable approach of model design, by visualizing which human parts a deep learning model focuses on. Based on an intuitive interpretation of model behaviors that lead to inaccurate results, specific improvement of model architecture is inspired. Our improved results suggest that making existing models explainable could effectively shed light on designing more robust models.