Pedestrian attribute recognition aims to predict the predefined attributes of a target pedestrian, generating a structured description of the pedestrian, which includes levels of 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 remains a hot topic. During the past decade, numerous methods have been proposed to facilitate PAR. In early studies, researchers usually take the whole pedestrian image as input, which leads to insufficient emphasize on some valuable information for fine-grained attributes. In order to solve this problem, some researchers propose to take the segmented local body parts as input. To more accurately locate the area associated with a specific attribute, various types of auxiliary modules such as attention mechanism and key points detection modules have been added. Thereafter, some researchers have suggested that mining information such as attribute relations can be of help to improving recognition accuracy. As a result, relationship matrix, context information and other image processing techniques are integrated into this task. However, current researchers are still faced with several challenges. Since the pedestrian images are captured in surveillance scenarios, resolutions of these samples can be rather unsatisfied and the captured identity can be seriously motion-blurred, resulting in poor image quality detrimental for a robust PAR. Also, the demand of PAR to be pedestrian orientation-invariant and time-series consistent make the task more complicated. Moreover, for samples captured by outdoor surveillance cameras, the image quality can be influenced by variation of illumination and weather as well. Considering the above challenges, this paper analyses the methods proposed in recent years. According to the type of input samples, pedestrian attribute recognition methods are categorized as frame-based ones, which independently judge on each frame of the input surveillance video; and video-based ones, which fuses multiple frames of the same identity to better explore the temporal consistence of pedestrian attributes within surveillance video. Basically, further improvement of PAR goes along two directions, one stream of research consider PAR as a stand-alone task and aim to foster it mainly from the supervision from the task itself. While the other views PAR systematically by considering its connection with other tasks intrinsically informative to PAR for the leveraging of further prior restrictions.This paper is organized as follows: the research background and the concept of pedestrian attribute recognition will be introduced at first. Then, pedestrian attribute recognition methods will be summarized and classified. After that, this paper will analyze and compare the current commonly employed datasets and experimental results on these datasets. Finally, we will discuss the future development direction of this field.