A Kinect-Based 3D Object Detection and Recognition System with Enhanced Depth Estimation Algorithm

In this paper, we present a system for 3D object detection and recognition with an enhanced depth estimation algorithm. The system is based on Microsoft Kinect and it uses state of the art Deep Neural Networks (DNN) for object detection and recognition. In addition, a robust algorithm for depth estimation was developed to overcome Kinect depth image accuracy limitations which are attributed to the presence of noisy pixel values, the large variation of pixel values for the same object and the inaccurate identification of objects bounding box. The proposed depth estimation algorithm uses statistical calculations (i.e. Mean, Median, etc.), to refine the depth image and remove or reduce the effect of the noisy pixels and other limitations. Experimental results validate the behavior and performance of the complete system and showed that the proposed depth estimation algorithm has increased the depth estimation accuracy to 88% from 82% in comparison with traditional algorithms (
$\approx 5\%$
accuracy enhancement).