FAFVTC: A Real-Time Network for Vehicle Tracking and Counting

In a complex traffic environment, the detection and association of moving objects can easily lead to tracking errors. This work proposes a novel attention mechanism called MCSA, which integrates multi-spectral attention and spatial attention. Additionally, a fast and anchor-free real-time vehicle tracking and counting model named FAFVTC is constructed. MCSA is used for extracting the features of moving objects, while FAFVTC is able to better detect and associate these objects. The effectiveness of the FAFVTC method is verified on the UA-DETRAC dataset. FAFVTC outperforms existing techniques with a 1.3 improvement in the PR-MOTA metric and a 2.16 improvement in the MOTA metric. The average tracking speed achieved is 27.9 FPS. The experimental results demonstrate that the proposed approach enables fast and accurate vehicle tracking and counting.