The Multi-Object Tracking (MOT) remains a crucial yet challenging task in artificial intelligence, especially in real-world applications. The difficulty stems from the large number of targets and their complex motion patterns. High-speed movements can cause significant frame-to-frame displacement, leading to tracking failures or incorrect target associations. Similarly, high acceleration can result in significant deviations from predicted positions, causing errors. To address these issues, we introduce BytetrackPro, an improved version of Bytetrack, designed with a more realistic motion model. BytetrackPro is robust against high-speed and high-acceleration targets, as well as camera shake. We extend the traditional Kalman filter by increasing the dimensionality of state estimations to incorporate more motion cues and introduce acceleration into the Kalman filter's updates for more accurate motion state estimation. Additionally, we replace the original covariance matrix with one that includes velocity and acceleration data. We also introduce the Motion Mahalanobis Distance (MMD) as a replacement for the traditional Intersection over Union (IoU) metric. BytetrackPro retains the low computational overhead of Bytetrack while delivering competitive performance improvements across datasets including MOT17, MOT20.
