The Research On Pedestrian Detection Based On Improved GMM And SVM

In recent years,computer vision and intelligent surveillance have been progressed rapidly.Automatic pedestrian detection has been widely applied to many fields,such as security,crowd control,etc.This thesis utilizes the way of pattern recognition to achieve the task of detection and uses a highly robust SVM classifier based on Histogram of Oriented Gradient.To reduce the computing complexity,it needs to detect the moving regions first.In addition,image pre-processing is also important in processes of feature extraction and moving regions detection.We perform some research about image filtering and edge detection.Edge detection undoubtedly plays a significant role in image processing.To improve the Canny operator,this research proposed an improved Canny operator which firstly utilized Anisotropy-based Gaussian filtering to reduce the effects of noises while it can protect the edge sharpness.And then it used OTSU-based Genetic Algorithm to search the optimal high and low thresholds automatically,which can keep away from manual setting.In the experiment of Lena,we gained the optimal thresholds(227,84)and the variance between classes was 3833.This approach effectively reduced the rate of false edges and improved the accuracy.In computer vision,it is crucial to obtain areas of interest.Traditional Gaussian Mixture Model providing a method of high-level modeling has had a good result,but when the environment suffers from changes of illumination and shadow,this method will detect wrongly.Therefore,this thesis proposed a new GMM based on gradient.Primarily,it used the Scharr operator to calculate the gradients for image pixels and applied these values(Gradients and Pixels)in GMM to get the backgrounds of video and then it was necessary to filter and connect image using morphological method.At the final stage,it did an AND operation for two kinds of backgrounds we gained.Our research compared results of experiments in three indoor and outdoor videos.Experiments showed that this approach is effective to eliminate the negative effects from illumination changes.This thesis utilized HOG and SVM to detect pedestrian in moving regions.For sets of people samples,we collected positive samples from opening MIT and INRIA databases and randomly produced negative samples from pictures without pedestrians using C++ and OpenCV.Then the method extracted HOG features from samples and trained an available SVM classifier.To further improve the performance of detection,it optimized the experimental results through least-square fitting to position the region of people and remove the error points.The experience indicated that the detecting rate of proposed method can reach at most 97.5%,so our approach has a relatively high precision and robustness in practical applications.