Appearance based pedestrians’ head pose and body orientation estimation using deep learning

摘要

Highlights•CNN is used as a building block to represent pedestrian head-pose and body-orientation classes with low-resolution images.•The proposed system is applicable to both still images and image sequences.•Two separate big datasets for head pose and body orientation are prepared to employ deep learning.•Only grayscale images from 2D cameras are considered as input to the proposed model.•Promising classification results are achieved, which are compared to current state-of-the-art approaches.AbstractPedestrian orientation recognition, including head and body directions, is a demanding task in human activity-recognition scenarios. While moving in one direction, a pedestrian may be focusing his visual attention in another direction. The analysis of such orientation estimation via computer-vision applications is sometimes desirable for automated pedestrian intention and behavior analysis. This paper highlights appearance-based pedestrian head-pose and full-body orientation prediction by employing a deep-learning mechanism. A supervised deep convolutional neural-network model is presented as a deep-learning building block for classification. Two separate datasets are prepared for head-pose and full-body orientation estimation. The proposed model is subsequently trained separately on the two prepared datasets with eight orientation bins. Testing of the proposed model is performed with publicly available datasets, as well as self-taken real-time image sequences. The experiments reveal mean accuracies of 0.91 for head-pose estimation and 0.92 for full-body orientation estimation. The performance results illustrate that the proposed approach effectively classifies head-poses and body orientations simultaneously in different setups. The comparison with existing state-of-the-art approaches demonstrates the effectiveness of the presented approach.