Multi-task Learning for Pedestrian Body Parts Detection and Multi-attribute Classification

Abstract

Pedestrian analysis plays a vital role in intelligent video surveillance and security-centric computer vision systems. Despite that the deep convolutional neural networks (DCNNs) achieved remarkable performance in computer vision, learning fine-grained features of pedestrian attribute tasks is a challenging task in complex surveillance scenarios. In this paper, we proposed a new DCNNs framework for pedestrian analysis, with the main idea of integrating different learning tasks of pedestrian body parts detection and pedestrian attribute classification, which called Hyper-pedestrian convolutional neural network (HP-CNN). Additionally, the proposed HP-CNN bring some advantages: 1) Squeeze-and-excitation block (SE-block) strengthens the representational power of networks by selectively emphasising informative features; 2) Multi-scale feature fusion concatenates more fine-grained information from both the low-level and high-level and enhances the contextual information from different convolutional layers. Experimental results conducted on the largest pedestrian attributes dataset show that the proposed HP-CNN obtained better pedestrian analysis results of both body parts detection and attribute classification, compared to the tested state-of-the-art methods. Especially, the auxiliary task of body parts detection can dramatically boost the performance of attribute classification.