According to the 2021 AI technology trend forecast from Gartner, AI technology has become a key technology in ADAS/ADS systems, which will be evolved from L3 to L4/L5 autonomous driving systems in the coming decade. The key technology required to enable autonomous driving system is the reliable object recognition and detection. The current AI-based object recognition and detection majorly focus on the camera data, which suffers from the low accuracy under low light and inclement weather conditions. Thus, it is required to combine the AI-based camera object detection with object detection of ToF sensors to ensure reliable and robust object detection. However, the existing object detection of ToF sensors still adopts traditional computer vision methods as well as digital signal processing methods, which suffers from limited detection accuracy and medium to high false alarm rate. On the other hand, the existing AI computing systems available in the market do not support the inference of AI models on 3D point cloud data from ToF sensors so that it still requires industrial PC with GPU card to meet the real-time processing, which suffers from high hardware cost and could not meet the production purpose. In order to overcome this problem, this project is entitled as “Deep Learning Accelerating System for AI Object Detection based on TOF Sensors,” which is planned as sub-project 3 to support the main project entitled as “Toward an Intelligent Vehicle Lidars without Moving Parts”. The goal of this project is to develop an efficient deep learning accelerating system that is able to support 2D CNN and 3D CNN models to enable the AI object detection based on 3D point cloud data captured from ToF sensors like solid-state lidar developed in the main project as well as the automotive grade mm-wave radars.
