A Research Of Video-based Pedestrian Detection And Tracking Algorithm

Video-based pedestrian detection and tracking is one of the key problems in computer vision which is widely used in human computer interaction, intelligent driving, and video surveillance. In recent years, it has become the frontier and hot topic in computer vision. Although the researchers made a lot of progress in this field, there are still many key issues remaining to be further researched for the pedestrian detection and tracking under complex scenes. This dissertation focuses on the above problems and covers the following works:1. In order to deal with the missing detection due to pedestrian deformations, we propose a new feature called deformable pyramid aggregate channel feature(DP-ACF) which is extracted from deformation area and pyramid pooling of channel feature. The DP-ACF feature can effectively reduce the missing rate and significantly improve the performance of the pedestrian detector. In addition, in order to deal with the false positive detection windows with small size around the pedestrian, we propose a new non-maximum suppression(NMS) algorithm called time scale non-maximum suppression(TS-NMS) which aggregates time and scale information to non-maximum suppression algorithm. The TS-NMS method can noticeably reduce the number of legacy windows which are false detected.2. When we track multiple pedestrians at the same time in a video, the trajectory regularly faces the problem of break and mismatch because pedestrian has been occluded. After analyzing the framework of online multiple objects tracking algorithm, we study a multi-object tracking algorithm based on the trajectory confidence and double-association, combined with the online-learning similarity model. We find that our method can successfully achieve even under occlusion and improve the accuracy of multi-object tracking.3. We design a complete pedestrian detection and tracking system and discuss the key steps in system design, including method for cropping training images, bootstrap rounds of training strategy and pedestrian detection and tracking evaluation. We implement an experiment platform for pedestrian detection and tracking based on Matlab GUI, whose functionalities include data labeling, performance and evaluation of detection and tracking algorithm.