Research On 3D Object Detection Method Based On Point Cloud

In recent years,with the continuous development of the autonomous driving field,traditional 2D object detection algorithms based on images can no longer meet the needs of this field.Therefore,3D object detection algorithms based on laser point clouds have received widespread attention,because laser point clouds well preserve the geometric information,depth features and other characteristics of objects in 3D scenes,enabling direct detection of objects in 3D scenes.To address the problem of the shape and scale variation of objects in images,this thesis conducts research on image feature extraction methods based on deformable convolutional operations to achieve adaptive feature extraction for image targets.To address the problem of image features lacking various scale information and spatial distribution information of targets and being incompatible with object detection tasks,this thesis conducts research on multiscale attention mechanisms,spatial attention,and channel attention to optimize image features.To address the problem of point cloud features lacking semantic information of the scene and image features lacking 3D spatial information,this thesis conducts research on cross-modal feature fusion methods in voxel space and achieves the fusion of 3D spatial information and semantic information of the scene,complementing each other to improve the detection results of 3D objects.To address the problem of limited multiscale information in fusion features and the high time complexity of relevant attention algorithms,this thesis conducts research on channel attention-based multiscale feature generation algorithms,separable self-attention algorithms,and deformable cross-attention algorithms to achieve multiscale information generation in fusion features and reduce the time complexity of relevant attention algorithms.Comparative experiments are conducted on the relevant technologies that affect the detection results of 3D objects.In various forms of cross-modal feature fusion comparative experiments,the average accuracy m AP on the nu Scenes test set reaches 61.2%,and NDS reaches 68.4%.Compared with other cross-modal feature fusion methods,the detection results are the best,verifying the effectiveness of the cross-modal feature fusion method based on voxel space.In the comparative experiments of different image feature extraction methods,when using deformable convolutional operations completely,the average accuracy on the nu Scenes test set reaches 63.8%,and NDS reaches 69.2%,which is better than other feature extraction methods,verifying the effectiveness of deformable convolutional operations in adaptively extracting various scale and shape target features.In various 3D object detection model performance comparative experiments,the detection model proposed in this thesis achieves an average accuracy m AP of 70.1% and NDS of 73.0% on the nu Scenes test set,which is better than the detection results of other object detection models.Observing the 3D object detection results in the images of the nu Scenes test set,the experimental results further illustrate that the detection model proposed in this thesis has robustness in low-light and crowded scenes.