Joint Pedestrian Detection and Attribute Recognition Feature Learning

In recent years, with the increase of video surveillance equipment, the amount of video data and complexity are growing. How to obtain structured data has become a problem. In surveillance videos, people extremely want to obtain valuable information about the position and characteristics of pedestrians, such as the pedestrian’s gender, age, appearance. Therefore, as a basic computer vision task, pedestrian attribute recognition plays an important role in the fields of image retrieval, pedestrian re-identification and video data structuring. However, the current method for obtaining pedestrian attribute information from images is to cascade the pedestrian detection algorithm and the pedestrian attribute recognition algorithm. This method requires a lot of time and the effect of the attribute recognition is affected by the detection results. To solve these problems, we design an integrated network that includes the pedestrian detection and the attribute recognition. Based on the integrated network, we improve the accuracy of detection and recognition by resetting anchor size, pre-training model and using GIOU loss. Experimental results show that after improving the integrated network for pedestrian detection and attribute recognition, the inference time of a single picture is 18ms under the condition of Nvidia Titan RTX graphics, which is equivalent to YOLOv3. Tested on the WIDER Attribute dataset, mAP 50 reaches 65.25% and mAP 75 reaches 34.13%.