A Method of Pedestrian Fine-grained Attribute Detection and Recognition

The detection and recognition of pedestrians is one of the hottest research issues. This paper presents a method that employs the fusion of convolutional neural network (CNN) models based on multi-task learning for multi-attributes, aiming to solve the problem of a low degree of accuracy in detecting and recognizing pedestrians fine-grained attributes under complex circumstances. First, the model implements double detection for data preprocessing. Specifically, this paper uses the CNN model twice to detect pedestrians under complex circumstances. In the first detection, it conducts the coarse-grained detection for the whole pedestrian; in the second detection, based on the first, it conducts fine-grained detection of and then recognizes pedestrian subcomponents. Further, to address the problem of a low rate of correct recognition for the fine-grained detection of the attributes of pedestrian subcomponents, we use the concept of the fusion of multi-task learning for multi-attributes, select the best recognized result for every attribute according to the fused results of different CNN classification models by the arrogance voting method, and then use a customized decision function to achieve a more accurate recognition of the coarse-grained attributes of pedestrians. Experimental results show that the proposed method achieved better performance than other recognition methods of pedestrians attributes.