Parallelized Acceleration Of Deformable Part Model And Its Applications

Object detection & recognition in video and images is one of the fundamental issues to be solved in compute vision field,and also a research challenge as well as a research hotspot in recent years.It has been widely used in video surveillance,humancomputer interaction,medical diagnosis and national defense industry.Nowadays,with the rapid popularization of vision sensors,various practical applications are in growing need of fast and high-performance object detection technology.Having experienced the development from artificial rules to data driven,object detection technology performs more and more compatibly on meeting to the practical application requirements.Deformable part model(DPM)is an excellent data driven approach for object detection.DPM algorithm has greatly improved the accuracy of object detection due to combining the root model with deformable part models.While the DPM algorithm has such many excellent properties,its detection speed is extremely slow.The average detection time per frame of DPM is more than ten seconds in Pascal VOC 2007 dataset,which becomes a bottleneck in its practical application.In order to overcome the speed drawback of DPM,we accelerated DPM with both the GPU(Graphics Processing Unit)parallel computing by hardware and algorithm redesign by software.GPU computing can fully exploit the parallelism of the DPM algorithm;DPM mixture model can effectively prune hypothesis in the feature pyramid,highly reduces the amount of calculation of DPM final model and improves the efficiency of detection process.After parallel acceleration,DPM algorithm not only performs excellently in Pascal VOC 2007 database,but is successfully applied in vehicle detection and license plate location.Parallelized DPM greatly improves the speed of vehicle and license plate detection without reducing detection accuracy,thus has basically reached the demands of practical application.In this paper,the main research contents and results are as following:1.Accelerating DPM algorithm with GPU.In order to break though the bottleneck of DPM detection speed,we accelerate DPM using GPU hardware,which is based on CUDA platform.We get more than 130 times speed up in Pascal VOC 2007 database with average detection rate of 10 FPS.2.Redesigning algorithm for parallel computing.Unlike traditional serial optimization with DPM algorithm,we propose parallelized optimization which is based on effective hypothesis pruning with mixture model.After algorithm redesigning,it is more convenient for parallel computing.We get more than 200 times speed up in Pascal VOC 2007 database with average detection rate of 16 FPS.3.Applying parallelized DPM in vehicle and license plate localization.We apply parallelized DPM algorithm in vehicle and license plate localization successfully,which already meets the demands of practical application.Both detection recall and precision are above 95%,the speed of vehicle detection is around 300 ms perimage and plate license was about 100 ms per frame.We design a complete auto license plate recognition system based on it,whose comprehensive recognition rate is around 90%.