Research On Pedestrian Detection And Tracking Based On HOG And Template Matching

Pedestrian detection is a key problem in machine vision. It has importantmeaning for improving the quality of life in contemporary society, and it is becominga research hot spot during recent years. For example, intelligent monitoring systemsfor buildings, vehicle auxiliary systems, motion analysis, advanced man-machineinterface application, these areas are popular application areas of pedestrian detection.Since the diversity of environmental detection background, illumination variation, theuncertainty of pedestrian movement and posture make pedestrian detection differentfrom other general target detection.Currently, the main detection methods of the pedestrian detection is based onhistogram of oriented gradients features and based on haar features. Detection methodproposed in this paper, is based on histogram of oriented gradients features combiningwith the template matching to improve the accuracy of pedestrian detection. Becauseof the uncertainty of pedestrian’s position, particularly the uncertainty of the fourlimbs, will greatly influence the detection results, and the pedestrian’s head shoulderposition has a good invariance regardless of how is the pedestrian’s position. So weadd the template of human head shoulder to do detection on the head shoulder part ofthe target.This method can enhance the detection rate. Then, this paper added trackingmodule after detecting pedestrian, using CamShift algorithm which has a goodreal-time performance combining with the particle filter algorithm to achievepedestrian tracking. This paper formed a complete set of pedestrian detection system,which implements including pedestrian detection in still pictures, also includingmoving pedestrian detection and tracking in a video stream.The main contents are as follows:1. The first part describes the significance of pedestrian detection and the currentstatus of research. This part mainly describes the background of the pedestriandetection, which is increasingly important in machine vision nowadays. And this partalso describes currently popular pedestrian detection methods based on motioncharacteristics, methods based on multi-part template matching, as well as methodsbased on machine learning, and this part also analyzes these methods coming with their respective strengths and weaknesses.2. This part elaborates the methods, nouns, expertise of pedestrian detection.Pedestrian detection based on machine learning can be divided into two parts, featureextraction and classification. In the feature extraction part there are many global andlocal features like wavelet features, histogram of oriented gradients features, haarfeatures etc. In the classification part nowadays we commonly use SVM (SupportVector machine) and Adaboost algorithm and other classification algorithms. Theearly part of this paper focuses on the support vector machine algorithm and theChamfer matching algorithm for template matching. The HOG algorithm is explainedin detail in the following parts.3. This part shows the pedestrian detection method proposed in this paper, themethod based on histogram of oriented gradients features combining with thetemplate matching. This section focuses on the characteristics of HOG, the pedestriandetection method based on HOG characteristics and its shortcomings, as well as therelevant content about template matching.And in the light of the above, we proposethis paper’s method for pedestrian detection, and clarify the advantages of this methodcompared with other methods.4. This part mainly gives the pedestrian tracking method which is proposed inthis paper. Combining the CamShift algorithm with particle filter algorithm to achievepedestrian tracking. Focusing on elaborating CamShift algorithm and particle filteralgorithm concepts and their advantages and disadvantages, and describing the effectof combining these two methods.5. The final part gives the data of the experiments based on the proposedpedestrian detection method in this paper, and makes comparisons with other relatedmethods coming with a detailed analysis.Experiments show that the proposed algorithm improves the accuracy ofpedestrian detection, and implements pedestrian detection and tracking in staticimages and video stream, forming a complete set of pedestrian detection system.