Pedestrian Tracking Algorithm On Traffic Videos

Object Tracking is one of the most fundamental topics in computer vision,and pedestrian tracking is one important branch in object tracking.Pedestrian tracking is widely applied to scene surveillance,human-machine interaction,and advanced vehicle technology.While the state-of-the-art tracking algorithms achieved great success,there are still some challenging problems to be solved.Firstly,it remains to be a tough task to develop a tracking algorithm with both accuracy and efficiency.Secondly,the ground truth is often given by a rectangular bounding box,which contains not only the target,but also the background pixels.The background pixels in the initial bounding box will mislead the appearance model of the target.Thirdly,most features for representing the target only use the gray scale information,and are not robust to pedestrians.The third problem is especially serious when the targets are pedestrians,which have colorful cloths on them,and change their pose and shape while walking.This thesis introduced the state-of-the-art techniques in visual tracking,and their disadvantages in pedestrian tracking.The major contributions of this work are highlighted as follows.Firstly,a novel saliency-based H-S histogram feature is proposed instead of the local binary pattern feature used in TLD,which improves the tracking robustness when the pedestrian dramatically changes the shape and pose.A saliency mask is involved to filter out background pixels which introduce noise to the pedestrian appearance model,so that the accuracy of the detection module in the tracking framework is improved.Secondly,ensemble color feature is further proposed to make full use of spatial relationship of body parts.It divides pedestrian region into 6 parts with the consideration of human's physiological structure,computing the color histogram in each part in color spaces of RGB,normRGB,HSV,and Lab.The spatial relationship encoded in this ensemble color feature is helpful to improve the accuracy of the pedestrian appearance model.As a result,the accuracy of the detection module in the tracking framework is further improved.Thirdly,a metric is established in this thesis to evaluate the robustness of pedestrian appearance model.To avoid time-consuming evaluation experiments over large dataset,it is proposed to pre-evaluate the robustness of a pedestrian appearance model using a small set of the pedestrian and background samples.Among a number of pedestrian appearance models,it can rapidly find the most robust one.The above proposed pedestrian tracking algorithms are evaluated on the Caltech Pedestrian Database.It is demonstrated that these algorithms achieve higher success rate and average overlap than the state-of-the-art methods.