AANet: Motorcycle ReID Using Multi-Atrous Convolution and Self-Attention Mechanisms

Recent advancements in ReID systems have primarily focused on individuals and four-wheeled vehicles to meet the rising demand for public safety and traffic management. However, the domain of motorcycle ReID has mainly been overlooked despite its significance in security applications. Extracting features from both motorcycles and their riders poses a unique challenge. In this paper, we present AANet, an innovative deep-learning model designed explicitly for motorcycle ReID. AANet incorporates multi-atrous convolution and self-attention mechanisms to extract local and global features, resulting in a robust embedding for the ReID task. We optimize the loss function and augmentations to address the specific challenges of motorcycle ReID. Our method achieves competitive accuracy and efficiency, surpassing other techniques, with a rank-l accuracy of 86.60 and a mean Average Precision of 88.32 on the Motorcycles ReIdentification dataset, the first extensive motorcycle ReID database collected from urban traffic cameras. The proposed AANet model enhances security measures through improved motorcycle surveillance and identification capabilities.