Target Reconstruction And Recognition Based On Acoustic Echo Feature

3D target reconstruction and recognition have irreplaceable significance for the extremely hot fields such as assisted driving,automatic driving,artificial intelligence,traffic safety and intelligent monitoring.It is also the core technology of blindness aid,UAV scanning and other fields,so whether this technology can develop and mature is related to whether many fields can break the bottleneck.At present,there are three methods for 3D target reconstruction and recognition,which are time of flight method,binocular vision method and structured light method.Using time-of-flight method,lidar is widely used for its fast scanning speed and long scanning distance.But lidar works poorly when scanning objects with high reflectivity in rain and fog.Therefore,a target reconstruction and recognition system based on acoustic echo features is designed in this paper.Firstly,the point cloud filtering algorithm is analyzed according to the characteristics of point cloud data.Analyze the design ideas and shortcomings of convolutional neural network and PointNet network.The deep learning framework is introduced and selected to provide the basis for the subsequent hardware system design and network design.Secondly,the software and hardware are combined to complete threedimensional point cloud acquisition.The hardware system mainly includes ultrasonic probe,steering gear and FPGA board.The software system designs the driver of each module and the scanning logic of the hardware system.In terms of software,according to the characteristics of the output data of the hardware system,the coordinate conversion algorithm is designed and transplanted to FPGA.The resolution was verified by test experiments.Fusion of two point cloud filtering algorithm design ideas.A more suitable distance filtering algorithm is designed and implemented in FPGA.Through the comparison experiment before and after filtering,it is found that the contour of the point cloud model is clearer after filtering.It provides a point cloud model with higher accuracy for subsequent classification and recognition.Finally,FK-PointNet was used to classify the point cloud model.Farthest Point Sampling(FPS)and K-Nearest Neighbor(KNN)are used to farthest point cloud.The point cloud data of divided multi-local areas is used as the input of FKPointNet network.The local features were processed by PointNet local feature extraction,dimensionality reduction after multi-scale feature fusion,and global feature extraction by double pooling.Get the global feature.After the improvement,the identification accuracy rate is the main evaluation parameter,and the distance and category are the classification criteria.The recognition rate of PointNet network and FK-PointNet network is analyzed respectively,and finally found through the experiment.Finally,through the experiment,it is found that the improved FK-PointNet network has better recognition effect than PointNet network.