| In recent years,with the rapid development of automobile industry,the occurrence of traffic accidents has increased year by year.In order to reduce traffic accidents,Advanced Driver Assistance System has received more attention.The core of the Advanced Driver Assistance Systems is environmental awareness.Sensors such as cameras,Lidar,millimeter wave radar,etc.are used to sense the environment around the road,detect vehicles in front of the road,and warn of upcoming collisions.Traditional ways for object detecting require artificial presetting and extraction of corresponding features,and there are major drawbacks such as low accuracy and weak anti-interference ability.However,despite its great advantages in terms of speed and cost,3D object detection based on monocular vision is still greatly challenging.Compared with solutions such as Li DAR and stereo vision,the accuracy of the monocular method is far from sufficient for ADAS applications.For example,when using the KITTI 3D object detection benchmark to detect the category of cars,the average accuracy of the state-of-the-art monocular vision algorithm is 63% lower than that of Li DAR-based algorithms.This paper proposed a monocular 3D object detection method.In the case where there is only one camera and the depth information is limited,we proposed a geometric locality preserving regularization term which can improve the accuracy of 3D object detection from 13.88% to 21.80%.The main contributions are as follows:(1)Propose a real-time monocular 3D object detection with deep neural network Mo Net3 D which can achieve an average positioning accuracy of 95.50%.(2)Propose a geometric locality preserving regularization term.Using the proposed regularization term to optimize the corresponding loss function greatly enhances the ability of the original network in 3D object detection and pose estimation and improves the corresponding accuracy.(3)This paper implemented an efficient system of automatic driving assistance based on Mo Net3 D.It can process video images at a speed of 27.85 frames per second for 3D object detection and localization. |
