The Tracking Of Pedestrians’ Status Based On The Analysis Of Far-view Video And Near-view Video

Intelligent transportation is a very popular topic in the field of computer vision, machine learning and transportation, which involves multiple cross disciplines and is very valuable in the academic research and application. The research background of the article is a project named "the recognition of pedestrians safety status in the mixed traffic environment", it aims at the robust and real-time tracking of moving object.The accuracy of tracking is limited by the change of object’s shape and speed, the variation of environment illumination, the complexity of background, the occlusion of objects and the shake of the camera. Although the field of multiple objects tracking has been focused for decades of years and many algorithms have been put forward, the stable, robust and fast tracking is still a challenge.In the paper, we try to build a tracking system which is based on the visual perception and combines the analysis of far-view videos and near-view videos to track pedestrians and vehicles. The far-view videos are captured from the overlook view and the field of view is wide so that many objects are contained and we can get the action of the objects and the space relationship among the objects. The disadvantage is the lack of details so that we can’t apply the fine algorithms in it. To the opposite, the near-view videos are captured from the height of a pedestrian, and contain reach details so that it is possible to apply the fine algorithms in it. The disadvantage is it is hard to analyze the space relationship among the objects. The major contributions of the paper are listed as follows:1. Tow improvements to the foreground extraction are proposed in the detection stage. The first one, we update the background model using an adaptive probability to speed up the update process and improve the accuracy. The second one, we modified the way to judge the overlap of adjacent segments to reduce the amount of searching and Improve the efficiency of the algorithm.2. In order to deal with the occlusion in the tracking stage, we detect the merging and splitting of object explicitly and on this basis we improved the adjacent location based algorithm. A fast multi-object tracking algorithm is proposed. Besides, we proposed an improvement to the multiple hypotheses tracking, so that the accuracy and efficiency are both improved.3. We collected the videos of the complex traffic environment in reality from a far-view and a near-view, explored the corresponding of the different views, and realized the combination of the process of the far-view videos and the one of the near-view videos in the tracking of the pedestrians and vehicles, which is proved robust and real-time. In details, to the far-view videos, we use the coarse algorithm to get the position and the speed of the objects, which is very fast. And then we deal with the near-view videos with the fine algorithm based on the result of the far-view videos process meanwhile the result is correct. In this way, the balance between the accuracy of the near-view videos and the efficiency of the far-view videos is build.