Pedestrian Detection Based On GPU Acceleration

This thesis proposes an algorithm for pedestrian detection which based on GPU acceleration to reduce the detection time and improve the detection rate.First,introducing theory about object detection and GPU parallel computing.Then,using GPU to caculate the GMM algorithm in parallel and proposing an algorithm for pedestrian detection based on improved feature.Finally,accelarating the pedestriap detection in parallel.The thesis studies the pedestrian detection on two ascept:background generation and feature extraction.The main works and innovations is as follows:(1)Designing and achieving pedestrian detection based on GMM algorithm to solve the problem about detection time and acheving the accelaration about the GMM algorithm.According to the pixels,assigning threads for the GMM algorithm.Acheieving the model innitialization,pixel judgement,model update and model sort in GPU space.While the data input and output about background generation are achieved in CPU space.The results of the experiments show that the algorithm can achieve efective pedestrian detection and reduce the detetion time.It has practical value.(2)This thesis proposes an algorithm for pedestrian detection which based on improved feature and GPU acceleration.First,the original images are processed by Canny operator and get the images with enhanced edge information.And processing images in three scales with the idea of multi-scales without scaling to reduce the interference of background and the deformation effect of unified standardized.Then,dividing each image into six regions to solve the occlusion problem among pedestrians,The regions are head,left arm,upper body,right arm,left leg,right leg,they are divided according to the characteristics of pedestrian action.After that,selecting SILTP feature instead of LBP feature to get better performances in the situation of low resolution and changing illumination.Extracting SILTP feature in parallel as the texture feature in GPU space to reduce the time of calculation.At the same time,calculating the gradient information and amplitude information of six regions in GPU space and weighting the value of gradient with the value of the distribution characteristics respectively.So,the improved HOG features with 180 dimensions are got.The dimensions are much lower compared with traditional HOG feature and the amount of calculation are reduced.Finally,concatenating the features extracted in three scales,including HOG feature and SILTP feature.Outputting all features to CPU space from GPU space and achieving pedestrian detection by the linear S VM clas sifier.The results of the experiments show that the features can express the information about pedestrian well and the proposed algorithm has strong robustness to the changes of illumination,the changes of environment and performs better in the aspect of occlusion among pedestrians.Obviously,the proposed algorithm can achieve effective and fast pedestrian detection.