Fast UAV- Image-based Person Re-Identification at the Edge

Person re-identification (re-id) has been widely used in various applications, such as security surveillance, the retail sector, and the military. Although companies that manufacture security surveillance equipment, such as Hikvision, Huawei, and NEC, have developed and deployed person re-ied systems in real life, this research topic still has several opening challenges, such as high memory usage and computation cost, processes at the edge and re-identification on aerial images. This problem is crucial for UAV-based person re-id applications in the military and security surveillance. This study proposes a deep person re-id model that can balance the tradeoff between accuracy and inference cost when implemented on the edge device and images captured from UAV. Although the mAP score of our proposed model is 2.48 % lower than that of the SOTA model, experimental results verify that our model-based inference cost is much better than the SOTA-model-based result. Indeed, our approach's RAM usage and processing time are 157 % and 528 % better than those of the Sota model, respectively. This result also points out that our proposed model meets the real-time requirement on the Jetson-Xavier-based edge device.