监控场景下基于单帧与视频数据的行人属性识别方法综述及展望

Pedestrian attribute recognition aims to predict the predefined attributes of a target pedestrian, generating a structured description of the pedestrian, which includes semantic information like age, gender, clothing, accessories and other levels of semantic information. Due to its wide application in the field of video surveillance and security, pedestrian attribute recognition has been widely concerned by researchers. With the rapid development of deep learning, researchers have proposed many methods to recognize pedestrian attributes in order to obtain more accurate results. In view of the challenges faced by this task in complex scenes, such as unclear surveillance scenes, pedestrian status change, occlusion, etc., this paper reviews frame-based and video-based pedestrian attribute recognition methods in surveillance scenario. First, the research background and the concept of pedestrian attribute recognition are introduced, and the problems and challenges faced by the current research are pointed out. The pedestrian attribute recognition methods are classified according to two different sample types of “single frame” and “sequential frames captured from video”. The newly proposed methods are summarized on the basis of techniques and ideas adopted in the attribute recognition process. Then the current commonly employed datasets and experimental results are analyzed. Finally, from the four aspects of state-guided pedestrian attribute recognition, tri-dimensional attribute, multi-task fusion and new data set construction, the future direction of this field is prospected.