Research On Micro-doppler Image Detection And 3D Pointcloud Object Detection Based On Deep Learning

With the development of deep learning methods,object detection algorithms have made great achievements.Radar sensors can realize all-weather invisible sensing of objects in a non-contact manner.Therefore,this paper proposes a micro-Doppler image detection algorithm and a 3D point cloud object detection algorithm by combining the advantages of radar sensors’ strong robustness with the advantages of deep learning-based object detection algorithm with high precision.The main work of this paper is as follows:(1)In the research of 2D object detection algorithm,we propose an improved Faster R-CNN micro-Doppler image detection algorithm based on millimeter-wave radar.First,the data captured by the millimeter-wave radar is converted into a micro-Doppler image,enabling non-contact perception of object distance and activity.Then,Faster R-CNN is used to detect the micro-Doppler image and output the category of the activity and time of the activity.Among them,we improve the recognition performance of the model by improving the feature extraction network,RPN and ROI of Faster R-CNN.The experimental results show that the mean average precision of the improved model on the test set and measured data is increased by11% and 9.1%,and the detection time is not increased.The anti-interference experiment shows that the improved model has good anti-interference performance.In addition,a set of experimental system is built,and the results show that the algorithm proposed in this paper has good real-time performance and accuracy in practical applications.(2)In the research of 3D object detection algorithm,based on lidar,an improved Complex-Yolo 3D point cloud object detection method is proposed.First,the point cloud data captured by the lidar is converted into an RGB image under the top view to realize all-weather and high-precision 3D environment perception;then the complex-Yolo algorithm is used to output the size information and coordinate information of the object.Among them,we added a multi-scale feature fusion network to the Complex-Yolo algorithm to improve the algorithm’s ability to detect small object.Moreover,we have also improved the backbone network and head network of the Complex-Yolo algorithm,which(through appeal improvement)improves the overall detection performance of the network.The detection performance on the KITTI dataset shows that the 3D object detection algorithm proposed in this thesis improves the mean average precision of 2D object detection and 3D object detection in top view by 22.7% and 20.8%,respectively,compared with the original algorithm.It has reached a relatively advanced level,and the improved algorithm is very advantageous compared to mainstream algorithms in terms of detection speed,reaching 30-40 FPS.In addition,an experimental system is built in this paper,and the results show that the average detection accuracy of 2D overhead target detection and 3D stereoscopic target detection is 57.5% and 51.5,and the frame rate is 38 FPS.