Research On Ship Target Detection And Tracking Algorithms Based On SSD

With the continuous development of the inland shipping industry,the inland river cargo throughput is increasing,and the risk of water traffic accidents is also rising.The application of the inland river monitoring system has received wide attention.As an important part of inland river monitoring system,ship detection and tracking algorithm obtains coordinate information from the monitoring image and transmits it to the tracking algorithm,which can track the detected ship target correctly and provide an effective basis for subsequent judgment of ship's behavior.The robustness of traditional ship detection methods to the actual complex inland river background environment is generally poor,and it is difficult to meet the needs of the actual scene.At present,convolutional neural network has a great advantage over traditional algorithms by virtue of its efficient feature extraction method,and has achieved amazing results in the field of image.In view of the problem of ship target detection on inland waterway,this paper made the ship annotated data set based on the real ship monitoring video of each maritime bureau,and increased the sample number by sample enhancement.In this data set,this paper trains and tests the target detection algorithms Faster-RCNN,YOLO and SSD based on convolutional neural network,and summarizes the performance indexes such as accuracy,recall rate and detection time of the three algorithms.In view of the shortcomings of the network model of SSD target detection algorithm,the performance of different models is compared,and the MobileNet network model is selected as the main framework of feature extraction.The priori frame is improved to make it more suitable for the sample space of ship data set.The average accuracy and detection time of the improved model are improved,which verifies the effectiveness of the improved method.Finally,the kernelized correlation filters tracking algorithm and data association algorithm are combined into ship tracking module,and the ship detection and tracking system is composed of the improved SSD target detection algorithm.The problem of missing detection caused by ship occlusion is improved.The test results show that the system can detect and track ship targets correctly,and has good accuracy and real-time performance.It can meet the needs of detection and tracking in the actual inland river monitoring system.