3D Objects Tracking By GPGPU-enhanced Particle Filter Algorithms

In recent years, with the development of the technology, more and more intelligent products appear in the manufacturing process and daily lives. Machine vision becomes more and more popular because it can reduce human interference in some cases. Object tracking is one of the key issues in the field of machine vision. Objects tracking methods have been wildly used in the fields of video surveillance, motion monitoring, robotics and so on. Particle filter is one of the promising methods and is largely studied by more and more researchers, but it is difficult to apply to real-time objects tracking system because of its high computation cost.People expect a higher image quality when they play games on computers or watch movies. So the GPUs are developed very fast. Many researchers focus on the migration of general algorithms to GPU to make them faster. And general purpose GPUs are of many research interests. Also, Kinect which is a camera made by Microsoft, becomes attractive because it can provide a lot of useful image data for researchers.In order to reduce the processing cost without sacrificing the tracking quality, this paper proposes a new method for real-time 3D objects tracking, using parallelized particle filter algorithms by a MapReduce architecture which is running on GPGPU. Our work is as follows. First, the basic theory of particle filter and its application in image tracking are discussed. Second, the researches on the GPU and the CUDA architecture are done to improve the efficiency of particle filtering. Third, Kinect and OpenNI has been used in our system. Then, a new 3D particle filter algorithm is proposed with GPGPU enhancement. Unlike the conventional 2D-based objects tracking, 3D objects tracking employs depth information provided by Kinect so that accuracy of tracking is largely improved. To reduce the high computational cost, the particle filer algorithm is implemented in GPU and the MapReduce architecture often used in big-data processing is applied. The proposed particle filter tracking scheme is realized on a GPGPU of CUDA5.5 and experiments are carried out. Experimental results show that the new scheme works in real-time and tracks robustly.In this paper, three techniques are proposed. They are, a) depth-information based self-adaptive tracking windows, b) parallelization on the transition, histogram computing and the likelihood calculation for particle filters run in GPGPU under a MapReduce architecture, c) XYZ-related weighted histogram.