Face/object recognition and tracking on IoT device with 3D camera

Abstract

Traditional face recognition systems often use RGB images as input for feature extraction and classification. However, with the gradually decreasing cost of depth sensors, RGB-Depth(D) images captured using low-cost sensors are becoming comparably easy to acquire. This project proposes a deep learning face recognition model for RGB-D images and deploys the developed model onto the proposed CUDA accelerated IoT platform. The proposed Local Binary Pattern (LBP)-Depth-guided attention model extracts features on RGB, depth and LBP images and utilizes feature-level fusion mechanism to guide the attention on RGB images. Compared with Depth-guided Attention, both quantitative and qualitative experiment results indicate that LBP-Depth-guided Attention has better focus on the important discriminative regions and achieved improved recognition accuracies under several challenging conditions, such as occlusion, pose variation and facial expression changes.