Research On Key Technology Of People Instance Search For Surveillance Video

With the widely use of surveillance cameras, the amount of surveillance videos are growing dramatically. It will be helpful to solve a case and investigate the cause of accident by analyzing these surveillance video data. However, due to huge amount of data and the complicated content of surveillance video, there are not any effective approaches to retrieval object automatic from surveillance videos. In general, in order to find suspicious vehicle or person, people have to spend a lot of time to browse huge surveillance videos. Besides, errors often happen due to artificial reasons which make the process efficiency extremely low. Under such condition, it becomes an important and hot research task to retrieval interest objects quickly and efficiently from surveillance videos.Instance search aims to retrieval objects which are an appearance of the query object under different illumination、viewpoint and scale from huge amount of videos or images. Instance search is different from general object retrieval and copy retrieval. The general object aims to retrieval object which is similar in semantic level and copy retrieval aims to retrieval image which is similar in visual level, while instance search retrievals object at instance level. People instance retrieval is a special field of instance search and it aims to retrieval all appearance of the query people in different time and place. People instance retrieval covers many technologies such as image processing, pattern recognition and computer vision which has a wide prospect in practical applications.In this thesis, we study in detail surveillance video retrieval technology nowadays and give a detailed analyzing on the challenge of people instance retrieval in surveillance videos. Besides, we also review the current situation of people retrieval in surveillance videos. On that basis, we propose a people instance retrieval framework for surveillance videos according to the characteristics of retrieving people from surveillance videos and provide some methods to key problems in this framework. The main research works and contributions of this dissertation can be summarized as the following:1. Propose a people detection method for surveillance video. This method first gives the regions that people will be present by motion detection. After that, part-based people detection method will be used to detect people in these regions instead of the whole frame. The proposed method can avoid detecting people in the whole frame scope, while it takes the advantages of part-based method that can work well to pose variety and occlusion.2. A human part segmentation method with a method to compute people attribute are proposed. Based on the analyzing of person position in detected blob, an improvement is made on current salient region detection algorithm in order to make person more salient. And Grabcut segmentation method is introduced to make a more precise segmentation result. Also, a people signature is proposed using the cloth color. Based on the characteristics of human vision and cloth color, we make a quantization to HSV color space. Then we use the dominant color which is extracted by computing color histogram of cloth as cloth color to represent people feature.3. Based on the above process, an attribute based people retrieval approach is proposed. First we make a retrieval by matching the attribute of given query with people in database. After this, a similarity is computed based on SIFT feature to rank the above retrieval results. Besides, we propose an extension retrieval method to meet the need of searching people without given query people image. In this retrieval, people first search people by providing the attribute of their expecting query people, then a query people image will be selected as a query instance from retrieved results.Besides, based on the research on key technologies of people instance search, we design and implement a people instance search system for surveillance videos.