Pedestrian attribute recognition based on multiple time steps attention

Abstract

Abstract Pedestrian Attribute Recognition (PAR) plays an important role in intelligent video surveillance. This paper tackles two severe challenges in it i.e., complex relations between images and attributes, and imbalanced distribution of pedestrian attributes. Specifically, a new multiple time steps attention mechanism is proposed to boost the modeling of the relations. Different from existing attention approaches that only focus on the current and previous time steps, it also exploits the knowledge of next time step. By adaptively capturing the knowledge of multiple time steps, more contextual knowledge is exploited. Meanwhile, to alleviate the challenge of imbalanced distribution of pedestrian attributes, a focal balance loss function is developed by increasing the cost of those attributes difficult to recognize. The proposed framework is dubbed as MTA-Net, which is demonstrated to be effective on two benchmark datasets, i.e., PETA and RAP.