Software Development Of Forward-Looking Sonar Detection And Recognition System Based On Embedded GPU

With the deepening of human research on the ocean,the demand for marine resources and concerns about marine security have made detecting various targets or resources in ocean become a major research hotspot.The sonar system,as a key means of ocean exploration,has received great attention.However,domestic research on forward-looking sonar detection systems is still in its infancy,with immature technology,high cost and poor environmental adaptability.Therefore,the research of a forward-looking sonar detection and recognition system with strong adaptability,good effect,and efficient recognition for underwater targets has good application value.Aiming at the above situation,this thesis develops a forward-looking sonar detection and recognition system based on embedded GPU.Taking the Tegra X1 embedded GPU platform of NVIDIA as the core,the system adopts a modular and object-oriented design concept.The software system is divided into upper computer and lower computer.The lower computer software obtains the sonar DFT data from FPGA through PCI-E,and uses CUDA to complete GPU-accelerated beamforming calculations to obtain sonar image data.The foreground data is extracted based on the ViBe algorithm,the target image is extracted using morphological methods,the feature vector of the target image is extracted using SIFT algorithm and BOW thought,and the target recognition for the sonar image is realized using the S VM algorithm.The multi-threaded cooperation is used to transmit the data to the upper computer through the network.The upper computer realizes data storage and corresponding display and playback functions,with a simple and easy-to-use user interaction interface.The system test results show that the forward-looking sonar detection and recognition system studied in this thesis can meet the real-time and efficient application requirements.The target extraction success rate of all categories reaches more than 95%,and the target recognition success rate of all categories reaches more than 93%.