Research On Key Technology Of Pedestrian Sequence Retrieval In Intelligent Surveillance

In recent years, with the growing number of surveillance equipment, the number of videos is also increasing in the squares, intersections and so on. These surveillance videos can provide useful clues and ways for police to solve criminal cases. Due to the complexity of surveillance video contents, traditional methods, such as manual browsing and target of interest searching, are not only time-consuming, laborious, inefficient, but also easy to miss the key information. To make up for the shortcoming of manual retrieval, automated approach to retrieve pedestrian is proposed. The intelligent pedestrian sequence retrieval technology is also one of the hot topics in the field of computer vision.Pedestrian sequence retrieval method based on pedestrian detection, extraction of pedestrian attribute and trajectory in intelligent surveillance covers a wide range. In this paper, several key techniques are studied in terms of pedestrian in surveillance videos, and the problems involved are researched deeply. The main work is as follows:Firstly, traditional pedestrian detection methods need to detect the whole image based on multi-scale sliding windows. It will be greatly waste of time if there is no pedestrian or few pedestrian in the image. And it is not robust to background interference and partial occlusion. Foreground regions are firstly detected by the use of moving information in surveillance video. Then pedestrian is detected accurately and rapidly by the method based on aggregated channel feature.Secondly, on the basis of pedestrian detection, clothing color attributes are extracted by the way of multi-frame voting. In consideration of the influence caused by illumination and device, this paper improves the existing method based on multi-scale illumination estimation, which employs local reflection statistics, and presents a multi-scale illumination estimation model to correct colors of the color-biased image in standard source. In addition, a hierarchical classifier based on multiple color space features is designed to accurately recognize corrected clothing colors which divided into "color type" and "neutral type", and then subdivide into color type.Finally, in order to solve the incomplete trajectory of the target block caused by target occlusion and pedestrian omission, five states, including moving target, static target, missing target, partial match and leaving target, are established about the pedestrian detected from adjacent image frames. Then, data association about target pedestrian is established between consecutive frames to match the target. When the target is missing or partial matched, the target status is updated real-timely. Complete target trajectory is extracted while target appears again to match target further. In addition, two methods of pedestrian sequence retrieval including text information retrieval and the instance image in the surveillance videos are proposed based on the research of key technologies. And a pedestrian sequence retrieval prototype system, which obtains good detection effect on the acquisition of video data in the campus, is designed.In addition, the technologies of pedestrian sequence retrieval are studied in this paper. The rapid detection of multiple objects in complex scenes, fast segmentation technology of pedestrian contours, trajectory extraction with similar color clothing and other attributes (trajectory, height, gender, or not to wear a hat, etc.) need to be researched and explored further.