Pedestrian attribute recognition based on multi-task deep learning and label correlation analysis

Pedestrian attribute recognition is an extremely challenging assignment because of continual appearance variations, background clutter, pedestrian occlusion, and diverse spatial distribution of unbalanced attributes. Thus, we propose a multi-task deep model for pedestrian attribute recognition, which aims at the various attributes of each pedestrian and the poor quality of pedestrian images. In this model, a deep convolutional network called Mask R-CNN is firstly adopted to obtain binary masks of pedestrian bodies. Second, multiply by the features obtained from different convolutional layers and corresponding binary masks to eliminate background interference from the extracted image features. Then, select the most suitable combination of feature maps for each attribute by a voting mechanism. Finally, employ a correlation coefficient and conditional probability-based label analysis algorithm to integrate prior knowledge into the proposed network. This model can not only reduce the effects of image background, but also avoid the contradiction between recognition results of different attributes by establishing correlations between them. Our experiments are conducted on two datasets (RAP and PETA) with a large number of pedestrian images. Experimental results show that this method is superior to other existing methods.