People Detection And Tracking Using RGB-D Data

Supported by collaboration project of ZJU-UTS Joint Research Center on Robotics, this thesis discusses some issues about building a robust people detection and tracking system. Detailed works and results are as following:(1) A people detection and tracking system based on RGB-D data has been designed and implemented. It consists of four modules:data collection, data pre-processing, people detection and feature extraction, people tracking. Based on standard message definition, a non-coupled system has been implemented. This system has been successfully applied to the subway people counting project.(2) A multiple sources fusion method used to detect multiple people based on RGB-D data has been designed and implemented. It takes advantage of the usual people detection method based on Histogram of Oriented Gradients (HOG) and Head-Shoulder-Signature (HSS). Dempster/Shafer rule is used to fuse the two probabilities. It deals with the need to reduce both the FN rate and the FP rate. The method can give out a more robust people detection result.(3) A descriptor based on RGB-D data has been proposed. It’s based on the RGB-D people detection method and uses both the RGB information and the depth information, named the descriptor as Hist-D.(4) A hybrid people motion model based on the features of people’s trajectories is introduced. Point Distribution Model (PDM) is used to analyze the trajectories of people and then the trajectories are clustered into three types:business people type, normal people type and tourist type. Then Brownian motion model is used to predict the movement of people whose trajectory type is tourist type and constant velocity motion model is used to predict the movement of people whose trajectory type is business people type and social force motion model is used to predict the movement of people whose trajectory type is normal people type.