This project provides an official implementation of our recent work on real-time multi-object tracking in videos. The previous works conduct object detection and tracking with two separate models so they are very slow. In contrast, we propose a one-stage solution which does detection and tracking with a single network by elegantly solving the alignment problem. The resulting approach achieves groundbreaking results in terms of both accuracy and speed: (1) it ranks first among all the trackers on the MOT challenges; (2) it is significantly faster than the previous state-of-the-arts. In addition, it scales gracefully to handle a large number of objects.
I believe this is taken directly from SORT/DeepSORT. You can check those repos and the corresponding papers for more information. But in general I think the values were just slightly tuned by the DeepSORT author and no one bothers to modify them when they use Kalman filter in other trackers.
why the init of kalman filter's std is: std = [ 2 self._std_weight_position measurement[3], 2 self._std_weight_position measurement[3], 1e-2, 2 self._std_weight_position measurement[3], 10 self._std_weight_velocity measurement[3], 10 self._std_weight_velocity measurement[3], 1e-5, 10 self._std_weight_velocity measurement[3]]
could you tell some reason or source ?