Online multi-object tracking by detection based on generative appearance models

摘要

This paper presents a robust online multiple object tracking (MOT) approach based on multiple features. Our approach is able to handle MOT problems, like long-term and heavy occlusions and close similarity between target appearance models. The proposed MOT algorithm is based on the concept of multi-feature fusion. It selects the best position of the tracked target by using a robust appearance model representation. The appearance model of a target is built with a color model, a sparse appearance model, a motion model and a spatial information model. In order to select the optimal candidate (detection response) of the target, we calculate a linear affinity function that integrates similarity scores coming from each feature. In our MOT system, we formulate the problem as a data association problem between a set of detections and a set of targets according to their joint probability values. The proposed method has been evaluated on public video sequences. Compared with the state-of-the-art, we demonstrate that our MOT framework achieves competitive results and is capable of handling several challenging problems.