An efficient target re-recognition method combining EfficientNet and MeMOT

With the wide application of deep learning techniques in computer vision, target re-recognition has become a popular research direction. Traditional target re-recognition methods face many challenges when dealing with target matching in complex scenes and multi-camera fields of view, such as view angle changes, lighting condition differences and occlusion. In order to better cope with these challenges, this paper proposes a novel target re-recognition method combining EfficientNet and MeMOT, referred to as EN-ME. EfficientNet, as a highly efficient deep neural network, is able to extract rich features from images; while MeMOT provides strong spatio-temporal memory support for target tracking. By combining the two, our approach not only accurately matches targets, but also enables efficient target tracking in real-time video streams. We conduct experiments on several publicly available datasets such as Market-1501, DukeMTMC-reID and CUHK03. The experimental results show that the EN-ME method performs better in terms of mAP and Rank-l accuracy. In addition, we analyze the robustness, computational efficiency and scalability of the model.