Object Detection in RGB-D Image Based on Two-stream SSD
Rong Wang, Jian Zhang, Yonghui Zhang · 2020 IEEE 3rd International Conference of Safe Production and Informatization (IICSPI) · 2020
Aiming at an unsatisfactory detection performance due to the limitation of the object detection method using color image or depth image alone, a two-stream SSD detection network combining color image and depth image is proposed, the depth map channel network is added on the basis of the original SSD detection network, the color information and the depth information are merged in different ways on the feature layer of the two-stream network, and the different features of the color map and the depth map are fully utilized, thereby effectively overcame the disadvantages of detection by single image, improved the accuracy of the object detection and the robustness of the environment. The experimental results on the SUN RGB-D dataset show that compared to the SSD detection network using color image alone and the SSD detection network using depth image alone, the accuracy of the proposed network is improved by 3.0 percentage points and 8.2 percentage points, respectively, and the detection rate on the GTX 1080 Ti is about 17FPS.