Person Reidentification In Surveillance Videos

Intelligent video surveillance is an emerging research field in computer vision,and is commonly used in public security,intelligent transportation and other fields.Pedestrian re-identification is one of the core functions of intelligent video surveillance system,it is of great academic significance and great application value to study the related algorithms and techniques.In order to realize pedestrian detection and recognition in surveillance scenes,this paper studied the motion detection algorithm,person detection algorithm and person feature extraction and re-identification algorithm.Based on these researches,a pedestrian detection and recognition system was constructed to realize intelligent analysis of the video.The following work is done during the process:First,the motion detection algorithm was studied and implemented.This paper studied and analyzed the ViBe background modeling algorithm,on which the motion foreground detection was done and the model parameter was modified to improve the algorithm.Afterwords,average filtering method was studied and implemented to have a post processing on foreground image,which can acquire the final motion area.The influence of its parameters on the effect was analyzed,which is the basis of consequent work.Next,the person detection algorithm was studied and implemented.HOG feature was extracted on pedestrian head-shoulder and body.According to the head-shoulder area,cascade Adaboost classifier was trained by cascade structure to test pedestrian head-shoulder.According to the pedestrian body area,DPM algorithm was used to train the DPM model to test pedestrian body.The combination of the head-shoulder detection and pedestrian body detection has realized the fast detection of pedestrians.Then,person feature extraction and re-identification algorithm was studied and implemented.Firstly,the dynamic thresholding method was used to realize color correction for images,which can avoide the influence of chromatic aberration on color feature extraction.Afterwards,the human body was divided into the areas of blouse and trousers.Color histogram was extracted from different regions as pedestrian features.Based on the analysis of the color distribution in monitoring scenario and the defects of the existing color quantization methods,this paper present a nonlinear color quantization method.This method can obtain the robust color histogram quantified by look-up table.The color matching was carried out according to the histogram intersection,and the effect of body color features was verified by experiments in re-identification process.Finally,this paper designed and built a set of pedestrian detection and re-identification system to test the above algorithms.