Research On 3D Object Detection Algorithm Based On Heterogeneous Data

With the development of autonomous driving technology,the safety of vehicles requires real-time perception of the surrounding environment.As an important part of sensing systems,3D object detection technology has received more and more attention.However,the traditional algorithms based on monocular vision and depth images are difficult to meet the needs of current perception systems.Therefore,this thesis aims at the problems in 3D target detection,and uses heterogeneous data to detect 3D targets in autonomous driving scenarios based on deep learning target detection algorithms,combining two types of heterogeneous data such as point clouds and image data.The thesis studies the object detection algorithm of point cloud data,the multi-view image object detection algorithm and the heterogeneous data object detection algorithm.The main research contents of this article include:1.This thesis proposes a point cloud object detection algorithm with embedded attention mechanism.Faced with the disorder and sparseness of point cloud data,voxels are used to process the point cloud data.In order to facilitate the feature extraction using a convolutional neural network,the obtained voxel features are transformed into pseudo image features.Attention modules are inserted into the subsequent feature extraction network to improve the model's feature expression ability.Then,analyze and compare the performance of different attention modules to build an attention mechanism that is more suitable for the network,the average detection accuracy is increased by 3% on the original basis.2.This thesis proposes a 3D object detection algorithm based on multi-view images.This thesis combines the top view,front view and RGB image data,and designs a onestage object detection algorithm.Aiming at the problem that small object is difficult to detect,this thesis adopts a feature extraction structure similar to the feature pyramid to obtain a full-scale feature map to improve the detection effect,which improve the average detection accuracy by 17%.In addition,for the imbalance of positive and negative sample types during the training of the one-stage target detection algorithm,the corresponding loss function is modified to improve the detection accuracy,which improve the average detection accuracy by 16%.3.This thesis proposes a 3D object detection algorithm based on heterogeneous data fusion.The point cloud data and image data are respectively extracted by different feature extraction methods to obtain their respective features,and then feature fusion is performed,and then sent to the detection network to obtain specific object positions.Compared with the monocular vision algorithm,the heterogeneous data method has greatly improved,but the performance of the algorithm still has a gap with the mainstream algorithms.Therefore,the regional point cloud segmentation method is adopted to further improve the detection accuracy,which increase the average accuracy of heterogeneous data algorithm detection by nearly 30%,and the feature fusion method is studied,and the fusion method is improved based on the point cloud characteristics.