Research On Pedestrian Clothing Recognition Algorithm In Video Surveillance

With the upcoming of Safe China and Safe city, video surveillances are applied in different areas, which bring convenience to management of security. But, it also brings huge time and cost consumption to search for target information when a sudden accident occurred, because video surveillance data is huge. Pedestrians are important targets in video surveillance, their appearance is critical to pedestrian recognition. Therefore, effective recognition the appearance of pedestrians will not only play an important role in pedestrian behavior analysis and video retrieval, but also improve efficiency of video surveillance staff. As a result, accurate recognition of pedestrians is particularly important, which includes recognition of clothing, bag, hair style, hat and so on. Clothing is the most obvious characteristic of appearance which directly affects correctness of particular pedestrian retrieval. Clothing recognition of pedestrians has two main aspects:clothing color recognition and clothing type recognition.This thesis contains five parts, and it is organized as follows.(1) Research background, significance and current situation of some advanced methods of clothing recognition are introduced first. Then, a brief description of main work and structure of this thesis are presented.(2) Two modules of classical cluster algorithms and feature extraction have been detailed introduced, including the detailed description of Kmeans, SLIC, Grabcut and the extraction methods of color feature, HOG feature, LBP feature.(3) A simple and efficient recognition method is presented to solve the recognition problem of clothing color in the surveillance video. Firstly, HOG algorithm and Grabcut algorithm are combined to partition the pedestrian in the frames of video surveillance automatically and accurately. Next, an appearance partition model is proposed to effectively segment pedestrian into the upper body and the lower body and further into many small patches. The KNN classification method is used to determine the color of each patch, and the clothing color is determined through color vote of all the patches. Finally, experiments are carried out in the collection of image data set from surveillance videos, which have verified the validity and practicability of this method. The experimental results show that the clothing color recognition rate is up to 88.9% by the proposed method in this chapter.(4) An efficient and feasible method is presented to analyze the clothing type of pedestrians. Firstly, pedestrian detection is used to obtain a foreground rectangle containing pedestrian. Then, Grabcut algorithm and SLIC algorithm are combined to obtain superpixel set of the body of pedestrian. Next, taking advantage of superpixel merging method, we merge those superpixels to accurately segment the clothing regions. After that, we extract LBP feature and HOG feature from clothing regions and use SVM classifier to obtain clothing type tags. Finally, experiments based on our image data set from surveillance videos have verified the validity of our method whose recognition rate up to 84.6%.(5) A conclusion of the whole thesis is made, and describes the future work.