GPNET: MONOCULAR 3D VEHICLE DETECTION BASED ON LIGHTWEIGHT WHEEL GROUNDING POINT DETECTION NETWORK

zizhang.wu

Keywords: applications in vision, audio, speech, natural language processing, robotics

TL;DR: Method for detecting vehicle 3D information based on fisheye camera with high efficiency

Abstract: We present a method to infer 3D location and orientation of vehicles on a single image. To tackle this problem, we optimize the mapping relation between the vehicle’s wheel grounding point on the image and the real location of the wheel in the 3D real world coordinate. Here we also integrate three task priors, including a ground plane constraint and vehicle wheel grounding point position, as well as a small projection error from the image to the ground plane. And a robust light network for grounding point detection in autopilot is proposed based on the vehicle and wheel detection result. In the light grounding point detection network, the DSNT key point regression method is used for balancing the speed of convergence and the accuracy of position, which has been proved more robust and accurate compared with the other key point detection methods. With more, the size of grounding point detection network is less than 1 MB, which can be executed quickly on the embedded environment. The code will be available soon.