Advanced Tracking Algorithm for Multiple Dynamic Vehicles using Attention Mechanisms and Kinematic Models

Efficient vehicle tracking in urban environments is crucial for intelligent transportation systems but faces challenges such as nonlinear motion, dynamic backgrounds, and frequent occlusions. To address these complexities, we propose the AM-Vehicle-Track system, designed under the tracking-by-detection paradigm. This system integrates the Lightweight Convolutional Block Attention Module (LCBAM), which enhances detection accuracy by refining feature maps through channel-wise and spatial attention, and the Noise-Adaptive Extended Kalman Filter (NSA-EKF), which dynamically adjusts observation noise based on detection confidence to handle nonlinear vehicle motion. Additionally, the Byte data association technique is employed to mitigate issues with low-confidence detections and reduce identity switches. Tested on the challenging VisDrone-MOT dataset, AM-Vehicle-Track achieves a mean object tracking accuracy (MOTA) of 42.2%, an identity F1 score (IDF1) of 51.2%, and processes at 72 frames per second (FPS), outperforming existing state-of-the-art systems. These results highlight its robustness and efficiency in real-time scenarios, making it suitable for intelligent transportation applications. By significantly improving detection accuracy, tracking precision, and scalability, AM-Vehicle-Track establishes a new benchmark in multi-object tracking technologies and demonstrates its potential to enhance the reliability and efficiency of vehicle tracking in complex urban environments.