Fast Object Detection, Positioning And Motion Analysis

Vision system of human beings is quite precise and complicated. People are trying to copy such a system with the function of perceive to the comput-er with the help of booming computer science and artificial intelligence, and then born the computer vision. The root task of computer vision is to perceive, and both visual information based object detection and motion analysis are the basis for intelligent systems to understand the world. In industry, more and more image sensors are taking place of the ordinary ones as they can catch huge information that no other can, urging a speedup of image detec-tion technology. And visual information based 3D human reconstruction and motion tracking is growing up in the field of virtual reality, film animation, intelligent surveillance, sports and medical treatment.Visual information is usually expressed in two or three or higher dimen-sions, and the huge data inside catalyzed both information explosion and the difficulty to analysis data in real time. Then comes the high performance computing, with its technology of speedup, and more recently GPU compu-ting and OpenCL, into the computer vision field.This thesis is established in designing a real-time object detection and motion analysis system. It first review recent researches about object detec-tion and human motion analysis, and then implements real-time detection al- gorithm for industry products and real-time human reconstruction and track-ing algorithms with the help of GPU and other parallel computing methods. And with those algorithms, this thesis established a real-time object detection and motion analysis system. The major research contents and results of this thesis are as follows.1) Proposed a high speed and template based object detection approach. It can detect multiple texture-less objects from the same scene in less than 100 ms using a coarse to fine search. This approach is designed for industry conditions.2) Proposed an approach combining human reconstruction and tracking under sparse views. It learns the state space online of the tracking algorithm and feed back tracking results to reconstruction approach and improves both.3) Accelerated human reconstruction and tracking algorithms using GPU, leading them to real-time. It accelerated the algorithms using multi-core CPU and AMD GPU, achieving a speedup of 400 and speed of 10 FPS.4) Established a real-time computer vision system.The well designed software system is coded in C++, brings real-time computing and system engineering to computer vision with the help of high quality capture and store devices. The part with GPU algorithms of the system has won the championship in the 2011 AMD China accelerated computing contest.