Research On 3D Object Detection Algorithms Based On Lidar And Camera

With the increasing number of vehicles,traffic accidents occur more frequently.The realization of automatic driving will greatly improve driving safety,and the efficiency of the entire transportation system,and save time for users.Environment perception is a very important part of automatic driving.In this task,the use of deep learning in 3D object detection is a hot research subject in recent years.In view of this situation,it is of great practical significance to explore a more accurate and efficient 3D object detection method.In order to obtain a more accurate and fast 3D object detection algorithm,this paper first analyzes and compares dozens of 3D object detection networks in recent years separately from network input,data preprocessing,and network structure.According to the experimental data provided by the KITTI dataset official website,the network using data from Lidars and cameras as input is generally superior to the network using only Lidar sensors.Based on the original experimental configuration,KITTI dataset and the same experimental hardware environment are adopted to retrain the three better effect networks--AVOD,MV3D and F-PointNet,which eliminates the effect of the experimental environment on the experimental results in the original experiment,highlighting the effect of the network structure itself.Secondly,this paper uses the nuScenes dataset with larger data volume and richer data features to retrain and verify the above three networks,which can provide more powerful experimental results.Through the training and validation of the three networks,it can be concluded that with the same experimental configuration and data set,in terms of accuracy and speed,AVOD works best,followed by F-PointNet,and MV3D performs poorly.Finally,to get a better network,this paper optimizes the loss function of the first phase of the network according to the specific network structure of AVOD.Inspired by the GIoU loss function,a 3D GIoU loss function is proposed in this paper,which is used to propose the non-directional region in the first-stage.The final experimental result shows that the 3D GIoU loss function proposed in this paper has a certain degree of improvement on the network accuracy.It can also be seen from the visualization result that the improved network can effectively improve the result of 3D object detection.In summary,the improved network in this paper has certain advantages in accuracy compared with the original network,and it is more feasible in the automatic driving scenarios.