Improved Kalman Filter and Matching Strategy for Multi-Object Tracking System

This paper proposes a method that can improve the performance of multi-object tracking by modifying the Kalman filter which is used to estimate the state of the objects in the video sequence and matching strategy to prevent objects obtained at the current frame in the video sequence from being associated with the wrong trajectory. The Kalman filter used in multi-object tracking is conventionally structured to use only bounding box location information as an observation system. To improve the accuracy of the filter estimation, we added velocity information derived from the difference of positional measurements to the observed data and introduced a process in which the observation noise covariance varies with the confidence score of the object detection. This enables tracking that is more nearly ground-truth. We also introduced gating as an improvement of the matching strategy. Gating is configured to compare the speeds that each track of the object and the matching candidate of the results of object detection with the Euclidean distance, and if this distance exceeds a threshold, the matching is canceled. The proposed method is introduced ByteTrack, which is the base of several object tracking methods reported in recent years. As a result, we achieved 90.1 MOTA, 83.7 IDF1, and 0.117 MOTP on the training set of MOT17.