Target Pedestrian Tracking For Robot Following Person

As the robots' application fields expand,person following becomes one of the important functions of many service robots,such as supermarket shopping robots,military auxiliary combat robots,and etc.In the task of robot following person,a key technique is the target pedestrian tracking,which is required to provide target's location for the robot.In addition,as a special case of object tracking,target pedestrian tracking is also widely used in other fields such as video surveillance and car safely driving.The application scenarios of existing target pedestrian tracking algorithms are relatively simple,such as applied for video surveillance scenarios where cameras are stationary,or applied on smoothly moving cars or wheeled robots,etc.However,existing algorithms are easy to drift when applied on legged robots because of the robot' s motion will cause the camera severely shaking.According to the development demands of per-son following for legged robots,this thesis proposes three kinds of target pedestrian tracking algorithms,based on findings in areas of pedestrian detection,person re-identification,object tracking and others.The proposed algorithms have been applied on a quadruped robot successfully.Details are as follows:1.A target pedestrian tracking algorithm that combines pedestrian detection and person re-identification is proposed.Firstly,the algorithm detects persons in images using pedestrian detector.Then person re-identification is applied to decide which one of candidates provided by pedestrian detector is the target.The framework of pedestrian detection is feature and classifier.It applies fast feature pyramids algorithm,yielding considerable speedups.With a random forest classifier,the detector is efficient and accurate.For person re-identification,template matching technique is applied.First,the weighted color histogram fea-tures are extracted.Then Bhattacharyya distances between candidates and templates are calculated,based on which the target is identified among candidates.The experimental results indicate that the proposed tracking algorithm can basically meet the demands of robotic person following,but serious frame losing exists meanwhile.2.A novel tracker based on compressed sensing and on-line Boosting is proposed and applied successfully to target pedestrian tracking.This tracker is an improved one of the compressive tracker.The main idea is to select distinguishable weak classifiers from multi-channel compressed features via Boosting to build a strong classifier.This strong classifier is robust to many challenges like background clutter,illumination variation,etc.Experiments on the public dataset indicate that with the framework of the proposed algorithm,the performance of origin compressive tracker is improved a lot and better than the state-of-the-art results.In the experiments on the quadruped robot,the rate of frame losing decreases a lot.But the precision of target's location is relatively low.3.A novel target pedestrian tracker based on particle filter with dual motion model and dual observation model is proposed.To solve the problem of abrupt and sudden movements caused by moving and shaking cameras,we construct a dual motion model that consists of two components:random walk and feature points track.In order to build a robust appearance model,we combine both class-specific and target-specific detectors in observation model.The likelihoods of particles being a pedestrian and being the tracked target are both exploited.This further enhances the robustness of the tracker.Experiments in the public dataset indicate that the proposed target pedestrian tracking algorithm is applicable to many scenarios and achieves state-of-the-art results.In the experiments on the quadruped robot,the rate of frame losing decreases further and the precision of target's location is promoted much.Target pedestrian tracking algorithm only provides target's position in images,while person-following robot needs 3-D physical position of target.Therefore,we apply the principle of binocular stereo vision to the three proposed algorithms and conduct experiments on a quadruped robot.The results indicate the effectiveness and practicability of these three algorithms.