Region-based convolutional neural networks for occluded person re-identification

Abstract

In a variety of applications, including intelligent surveillance systems, targeted tracking, and assistive human-following robots, the ability to accurately identify individuals even when they are partially obscured is imperative. Such Continuous person tracking is complicated by the close similarity between the appearance of people and target occlusions. This study addresses this significant challenge by proposing a two-step, detection-first approach that uses a region-based convolutional neural network (R-CNN) as the re-identification (re-ID)solution. The model is specifically trained to detect occluded persons at different levels of occlusion before forwarding the image for the re-ID process. Three occluded-specific datasets are selected to evaluate the model's effectiveness in detecting occluded people. There are 379 distinct people in total, and each has five images obstructed from different angles. A sample of the data is taken to simulate various environment settings, and new data points are generated with different degrees of occlusion to assess how well the model performs under varying levels of obstruction. The findings demonstrate that the proposed person re-ID model is reliable in most circumstances, correctly re-identifying at 74% (Rank-1) and 90% (Rank-5). Although there is a decrease in accuracy as the number of distinctive people in the dataset increases, this does not significantly impact the tracking performance in various applications, which are expected to recognize a single person or a small group of individuals. Future works will explore refining similarity matching algorithms by delving into robust image comparison techniques, thereby addressing the challenges presented by occlusions. A critical aspect is to assess the model under diverse lighting conditions and investigate scenarios with multiple individuals in a frame. It is also beneficial to exploit high-resolution datasets, such as DukeMTMC-reID, and integrate finer contextual details, like clothing or carried objects. These collective efforts are essential for optimizing the model’s efficacy in practical applications and advancing person re-ID technologies.