3D Object Detection And 6D Pose Real-Time Estimation Based On Vision

Visual perception has been widely used in intelligent driving scenarios,and vision-based 3D object detection and 6D pose estimation algorithms can obtain more target information without increasing the cost of perception,but there are relatively few studies on this problem,and the performance of existing algorithms is difficult to meet the needs of vehicle intelligent driving.In order to realize the application of 3D target detection and6 D pose estimation algorithm to unmanned driving scenarios and improve the performance of algorithms,this dissertation makes full use of its sensor equipment to carry out research on 3D target visual perception in unmanned driving scenes,improves the existing perception algorithms,and optimizes its performance.The main research contents of this dissertation are as follows:(1)The 3D target detection and 6D pose estimation algorithms were expounded.The research status and development trends of 3D target detection technology and 6D pose estimation technology at home and abroad are summarized,and the problems of vision-based 3D target detection and 6D pose estimation algorithm applied to unmanned forklift platform are analyzed,and the significance of this study is pointed out.(2)Research on deep learning algorithms for object detection and P-n-P algorithms.Aiming at the problem of 3D target detection,the difference between 2D target detection and 3D target detection is analyzed,and the existing deep learning algorithm is improved to obtain 3D target bounding box on RGB images.Aiming at the problem of 6D pose estimation,according to the basic theory of camera imaging principle,the conversion relationship of each coordinate system in the camera imaging process and the P-n-P algorithm to solve the pose principle are derived,and the P-n-P algorithm is used to obtain6 D pose information from the 3D bounding frame.(3)Test the object detection algorithm using a homemade dataset.Build an unmanned forklift platform,collect data,self-make datasets,and test algorithms using open source datasets and self-made datasets.This dissertation introduces the composition of the dataset and the role of each part in the process of model training,algorithm testing,and verification.Using the open source dataset,the performance of the algorithm is tested,and the test results show that the detection speed of the algorithm can reach 45-50 FPS to meet the needs of real-time detection,but the detection accuracy is about 80-93% and the detection error is about 4.0-5.2 pixels.According to the requirements of data collection and real vehicle testing,build an unmanned forklift platform,use the platform for data collection,use the self-made dataset according to the algorithm requirements and the open source dataset format,and test the algorithm again using the self-made data set,and the results are similar to the test results under the open source dataset.(4)The improved algorithm of this dissertation is verified by using the unmanned forklift platform.The original algorithm is optimized,and the detection accuracy is optimized on the basis of ensuring the real-time performance of the algorithm.In this dissertation,the network structure and loss function of the optimization algorithm are used to achieve the optimization goal,the network structure part deepens the number of network layers,uses the convolutional layer with step 2 instead of the pooling layer,and introduces the residual network structure,and the loss function part distinguishes the nine key points into center points and corner points and redistributes the weights.The improved algorithm is tested with a translation rotation accuracy of about 99%,an average translation error of about 1.5cm,an average angle error of about 1.7°,and a detection speed of about 30 FPS.Through multiple experimental verifications,the optimization scheme proposed in this dissertation can significantly improve the detection accuracy of the original algorithm after sacrificing part of the detection speed.