Boosting Multi-Object Tracking Performance with Advanced Attention Mechanism based on Transformer

The Transformer architecture has been successful for a variety of tasks in natural language processing and has recently gained popularity in computer vision applications such as medical image analysis, traffic light monitoring, surveillance, and object tracking. This is due to the self-attention operation in Transformer, which enables global interactions between image patches, resulting in powerful modeling capabilities. However, the quadratic complexity in time and memory associated with the Transformer poses a challenge. To address this issue, we propose an advanced attention mechanism based on the Transformer architecture for multi-object tracking. The advanced attention mechanism consists of Transposed Self-Attention (TSA) and Cross Patch Interaction (CPI) modules. The TSA module utilizes a transposed self-attention mechanism instead of the conventional self-attention, operating across feature channels rather than image patches. It has a linear computational complexity with respect to the number of patches, resulting in reduced computational costs. Meanwhile, the CPI module is added following the TSA module to enable communication across patches, thereby enhancing the model's learning efficiency. This approach can efficiently process high-resolution images, achieving state-of-the-art performance in multi-object tracking with a MOTA of 72.8% on MOT17.