People Detection And Tracking For Service Robot

As humans have gradually stepped into the information society from industrial society, information and intelligibility products walk up to all aspects of people’s life. The aged tendency of population and the rising cost of manpower make the social demand for service robots become more and more urgent. People detection and tracking, as a basic function of service robots, play an indispensable role in human-computer interaction. However, due to the complexity and uncertainty of environment, the change of light and occlusion between people and so on issues, service robots still face many challenges in the detection and tracking of people.The research purpose of this paper is to design and implement a people detection and tracking system with good real-time performance and high robustness for service robots, thus make it as a basic function to provide support for the realization of human-computer interaction for service robots. Eventually, the service robots can perform various tasks better. To achieve this goal, this paper makes a detailed exposition and analyzes the various technologies and their difficulties. And on this basis, the following work has been done:(1) In view of one leg close to the ground with stability, in this paper, the laser sensor is installed at the bottom of the turtlebot robot to transform it. And then turtlebot robot can detect people through identification of human leg by laser. In order to improve the accuracy of the identification of legs, we put forward 10 features to describe the laser segment by analyzing the properties of human leg and use Adaboost algorithm realize the classification of human legs in this paper. On two data sets published online, our experimental results show that the method developed in this paper has high accuracy and real-time performance.(2) A people detection algorithm is proposed to deal with the far and near two kinds of cases according to the distance between the people and the robot by using rich visual information and depth information of Kinect sensor. The pointcloud pretreatment in the initial stage, including pointcloud sparse, ground detection, pointcloud classification and pointcloud clustering, make the scope of the people detection significantly narrowed down. For people from a relatively close distance, pointcloud clusters are projected into the depth image space, and then the template matching method is used to match the people’s upper body for people detection. For the more distant people, pointcloud clusters are projected into the RGB image space, using GPU to accelerate the extraction of HOG features, and the SVM is adopted for people detection and classification. In the final stage of testing, we use the nearest neighbor algorithm to fuse the detection results of laser and the detection results of Kinect and determine the location of people in space. Therefore, the accuracy of detection is improved.(3) In the phase of people tracking, a layered particle filter algorithm is proposed. MCMC particle filter algorithm is used in coarse grain space, and SIR particle filter algorithm is used in fine grain space. SIR particle filter overcomes the problem of particle degradation in particle filter and speed is very fast, therefore, people can be quickly and accurately tracked in fine grain space. MCMC particle filter algorithm as a complement to SIR particle filter solves the particle diversity loss and tracking failure when the number of targets changes of SIR particle filter. However, the speed of MCMC particle filter algorithm is low. and it should be used in coarse-grained space.(4) Flood tracking model is put forward based on the depth map. We implanted seeds in the initial position of people, and spread from the seed position until all of the points that contain the people enter the queue in the next frame, and then we implanted the tracked people center position as a seed, followed by recycling. This method is simple, and the running speed is very fast.