Multi-object Tracking In The Occluded Area Based On Optical Flows

Video target tracking technology is one of the hot issues in the field of computer vision and has a broad application prospect and huge economic value in military and civil such as military guidance, intelligent transportation, security surveillance. By analyzing the appearance and motion characteristics of individual goals detected in every frame, academics get the trajectory of moving object and achieve its behavior understanding and real-time tracking as well. Tracking of multiple partially occluded humans is a difficult problem, the key to solve this question is to establish a proper target model, usually the established model based on appearance and edge patterns easily influenced by dress color and light conditions may become inaccurate during the occlusion. The motion parameters of moving object can be estimated by optical flow method can get. According to the optical flow field of the same target which has globally consistent, and different targets which have different optical flow speeds and directions, this paper establishes the target model based on optical flow field.For the question of incorrect target tracking when a dynamic occlusion occurred between two or more objects within a surveillance video, this paper raises use the optical flow to track pedestrians in partial occlusion and interaction area,the contents are as follows:1) Getting the optical flow directions and velocity of pixels in foreground moving area by Lucas-Kanade optical flow method.2) Establishing a single target tracking model based on optical flow field through merging the optical flow directions by normalized simplification model, calculating the optical flow directions energy histogram, establishing a target model with the learning energy results of main direction and adopting the Camshift method process to realize object tracking.3) Multiple partially occluded humans tracking based on single tracking model. There are two cases of multiple targets motion state in the overlapped areas:(1) If the directions of moving objects are too different, the main direction can be used as constraints, and the occluded target matching conditions must be updated real-time;(2)When the directions are similar, the position of target is predicted by its optical flow speeds and directions, using the combination of this two cases can provide real-time multiple objects tracking.In this paper, we have realized multiple pedestrians tracking in cross-block area, and reduced the impact of pedestrians clothing color and video background factors. Experiment results show that the proposed algorithm has robust tracking performance overall.