A Novel Convolutional Neural Network for Head Detection and Pose Estimation in Complex Environments from Single-Depth Images

Computer vision based on neural networks is an important part of modern cognitive research. As important applications, head detection and pose estimation have made breakthrough progress in recent years. Compared to RGB sensors, depth cameras can provide a reliable solution in unstable or poor lighting conditions. An efficient pose estimation method relies on accurate head centre localization. Based only on depth images, a new convolutional neural network named HDPNet, which implemented complete head detection and pose estimation in complex environments, was proposed. For the head detection part, HDPNet adopted a convolutional neural classification network and the mean shift algorithm to achieve high-precision head centre localization, and for the pose estimation part, a novel guidance network with L2-norm was introduced to constrain the regression process of pose features. Moreover, soft label was adopted to encode the probability distribution between the pose ranges. To verify the performance of HDPNet, a series of experiments were conducted on four challenging public datasets: Watch-n-patch, the Biwi Head Pose dataset, Pandora and ICT-3DHP. Based on our experimental results with a comparison to state-of-the-art methods, the IoU of the head localization was improved by 2.2