OHMA: An Edge-Based Lightweight Occluded Target Re-Identification Framework for Exploring Abundant Feature Expression

Abstract

The rise of the Internet of Things (IoT) and the Internet of Vehicles (IoV) has accelerated the realization of smart cities, where cameras as interconnected consumer electronics (CE) are deployed across cities to capture target images. The widespread deployment of monitoring equipment has prompted us to focus on the target re-identification (Re-ID) issue. One major challenge about this issue is that the identified targets are often obscured by different obstacles, which leads to bad performance. In practical applications, the occluded Re-ID task is very significant to complete. Previous approaches have focused on improving the occluded Re-ID performance but have neglected the lightweight problem, which makes the model difficult to deploy in the real world. Therefore, this paper proposes a lightweight framework that ensures occluded Re-ID performance and deploys at the edge to solve the problem of long transmission time and high latency caused by wireless and cloud technology in CE. This framework tackles occluded target Re-ID issues by integrating omni-scale features with human keypoint estimation and multi-head attention mechanism (OHMA). To solve the vehicle Re-ID problem, we use the cutout method to simulate an occlusion scene due to the lack of occluded vehicle data. Then, The multi-head attention mechanism combines with the omni-scale network (OSNet) to learn vehicles subtle features. To deal with occluded pedestrians, human keypoint estimation focuses on non-occluded areas of pedestrian images by paying attention to visible information about the human body. The generated heatmaps fuse omni-scale feature maps to explore better feature representations. In addition, the HUAWEI Atlas 200I DK A2 is used to simulate real edge devices and evaluate the experiments on both public and real-world private datasets. The results demonstrate that our framework improves the occluded Re-ID performance while ensuring lightweight. Compared with the previous methods, OHMA displays advantages in occlusion scenes.