Research On 3D Object Detection Algorithm Of Unmanned Vehicle Based On Improved F-ConvNet Algorithm

3D object detection can not only output the center position of the target,but also output the length,width,height and rotation angle of the object.Therefore,it is widely used in the research of automatic driving environment perception.Among them,the F-ConvNet algorithm based on image and point cloud fusion uses mature 2D object detection algorithm to guide 3D object detection,which reduces the difficulty of learning and has good accuracy.However,it also has problems such as relying on 2D object detection results and inconsistent scale features.In order to further improve the accuracy of the object detection algorithm and have a better practical application in unmanned vehicles,this paper takes the F-ConvNet algorithm as the benchmark to improve and optimize it,and makes the final practical verification on unmanned vehicle.The main research contents and contributions of this paper are as follows:(1)Aiming at the problem that the performance of the F-ConvNet algorithm is limited by the accuracy of the 2D object detection results,the2 D object detection algorithm YOLOX with higher accuracy is selected as the detection framework,and its network structure is optimized as follows:the coordinate attention mechanism is added to the sampling places on the output features and feature pyramids of the backbone network,so that the network pays more attention to the region of interest;in the feature pyramid structure,some additional channels are added to the original feature layer to the output node for bidirectional feature fusion,so that the network incorporate more features.The experimental results show that adding coordinate attention mechanism and bidirectional feature fusion structure can improve the accuracy of YOLOX algorithm detection results.(2)Aiming at the problem of inconsistent scale features in the fully convolutional network structure of the F-ConvNet algorithm,an adaptive spatial feature fusion module is proposed.Assign a weight coefficient to the feature map of each layer,and retain more useful information by adaptively learning the importance of features at different levels,thereby suppressing inconsistent features,and introducing a Bin-based loss function into the network training process,which enables more accurate regression prediction.The effectiveness of the adaptive spatial feature fusion module and the Bin-based loss function is proved in both qualitative and quantitative analysis in the experimental link.The experiments on the real vehicle also verify the practicability of the improved algorithm.(3)Aiming at the problem that the inaccurate synchronization of the camera and the lidar will affect the accuracy of the detection results of the subsequent algorithm,a time-based soft synchronization method is used for time synchronization.Firstly,the Chrony and Gpsd programs under Linux are used to unify the system time of the industrial computer and the GPS time,then the PTP synchronization protocol is used to synchronize the time of the industrial computer and the sensor,and finally the data information of the two sensors is matched by the nearest neighbor frame matching algorithm.The experimental results show that the method has smaller synchronization error and can meet the processing needs of subsequent algorithms.