A Cost Matrix Optimization Method Based on Spatial Constraints under Hungarian Algorithm

At present, with the continuous development of single object tracking(SOT) tracker, more and more SOT tracker are applied to multi-object tracking(MOT) tasks. However, in the traditional method of affinity computation, affinity model is used as the metric of Hungarian algorithm, which has the low discrimination rate of similar objects and leads to ID switch easily. In order to solve this problem, we propose a cost matrix optimization method based on spatial constraints under Hungarian algorithm. In the data association stage, Kalman filter is used to estimate the motion vector of the object, so that the current position of the object can be linearly predicted. The weight of the cost matrix is modified according to the spatial relationship between the estimated position and the detection results, which is used for the subsequent re-identification task. It is worth noting that the above methods do not need extra training and can be directly used in other multi-object tracking models. Our method has been evaluated on MOT16(46.7%), MOT17(49.7%), and achieved the effect of SOTA. The entire results can be found on MOTChallenge website1 .