Research On 3D Target Detection Algorithm Of Autopilot Based On Monocular Image

With the rapid development of deep learning,artificial intelligence technology has been widely applied in various fields,the most prominent of which is the progress made in the fields of autonomous driving and intelligent robots.With the gradual increase in the number of motor vehicles,autonomous driving technology is receiving increasing social attention.The auto drive system is mainly composed of three parts: target perception,planning decision-making and control.Target perception is a key basic link of automatic driving.Only by accurately predicting and positioning the position of external three-dimensional objects such as vehicles and pedestrians,vehicles can avoid obstacles and ensure driving safety.The3 D object detection method based on monocular images has important research significance and broad application prospects due to its advantages such as low cost,easy installation and maintenance.Aiming at the problem of low detection accuracy in monocular images due to the lack of depth information of the detected target,this thesis conducts in-depth research on relevant algorithm theories,proposes improved algorithms,and conducts experimental verification analysis.The main contents are as follows:(1)The target depth is predicted using a geometric projection method and a depth neural network direct regression method,and the uncertainty of the prediction result is given.Then,the corresponding score is obtained from the uncertainty.At the same time,a depth fusion module is designed to fuse the depth predicted by the two methods according to the score,and finally,a balance adjustment parameter α is introduced,adjust the weight ratio of the depth obtained by the two methods in the final fusion depth to reduce the error caused by the different detection accuracy when the two methods predict separately.Through deep fusion,the accuracy of depth prediction is improved,effectively improving the accuracy of 3D target detection based on monocular images in autonomous driving scenarios.(2)Aiming at the occlusion problem of detected objects in images and the difficulty of detecting distant objects,a monocular 3D object detection algorithm based on attention mechanism is proposed.By adding the attention mechanism module CBAM(Convolutional Block Attention Module)to the algorithm model,and improving the channel attention module and spatial attention module of CBAM.By removing the maximum pooling and average pooling operations in the channel attention module and spatial attention module in CBAM,feature information is retained to amplify cross latitude interactions,thereby improving the feature extraction ability of the target and improving the detection accuracy of the target.Experimental results on the KITTI dataset show that this algorithm is effective and feasible in improving the accuracy of target detection.