Research On Detection Methods Of Point Cloud Vehicles Based On Deep Learning

Due to the continuous improvement of performance and stability of 3D data acquisition equipment such as Li DAR,and the increasing research and application in robotics,autonomous driving,artificial intelligence,and other fields,point cloud has become one of the most important forms of 3D data.As the preferred technology for tasks such as classification,segmentation,and detection in artificial intelligence,applying deep learning to point cloud data is also one of the research hotspots in recent years.The thesis takes the point cloud vehicles detection method based on deep learning as the research topic and focuses on how to apply the popular deep learning target detection framework to the unique point cloud data format.The main research content is divided into three parts.（1）Construct a point cloud vehicles dataset.Firstly,the collection equipment and collection methods used to construct point cloud datasets for vehicles detection tasks are introduced.The thesis adopts a semi-solid prism scanning Li DAR,which improves the point cloud density and resolution of the collected data.Then,the method of making and labeling the dataset and the method of cleaning the dataset is introduced.For 3D point cloud data,due to its special spatial structure,fine annotation can greatly promote the learning of the network,so the collected data is carefully annotated with the help of ROS tools.Finally,the related evaluation indexes of the 3D object detection algorithm are studied.（2）Research on deep learning point cloud vehicles detection based on 3D candidate box proposal and Io U loss.The thesis constructs a deep learning point cloud vehicles detector based on 3D candidate box proposals and Io U loss for detecting 3D vehicle objects from raw point clouds.The network learns in two parts.The first part of the network,in a bottom-up manner,extracts point cloud features through a point cloud codec,and then generates a 3D candidate box directly from the point cloud through a foreground point segmentation network and a 3D bounding box generation network based on interval loss.The second part of the network corrects the candidate boxes in the canonical coordinate system by jointly learning the point cloud features and the point cloud local spatial features.（3）Research on deep learning point cloud vehicles detection based on local feature perception.In order to enable the network to learn local features in the same sparse 3D point cloud data structure as the 2D convolution network,the thesis uses the 3D sparse convolution method to construct a deep learning point cloud vehicle targets detection network based on local feature perception.In the first stage,point cloud features are first extracted with an encoder-decoder network constructed with 3D sparse convolutions and then learned to estimate accurate intra-object local locations by using intra-object local location labels and foreground point labels from ground-truth 3D bounding boxes.At the same time,according to the attributes of the vehicles,the generation strategy of the 3D candidate frame is improved from the bottom-up anchor-free scheme to the anchor-based scheme to better learn the vehicle point cloud features.Then,the local location features within the predicted object are pooled by a region-of-interest-aware-based point cloud feature pooling method.In the second stage,the local position information is aggregated so that the network can better capture the local geometric information of the object to accurately evaluate the confidence of the candidate boxes and correct their positions.