Study On The Method Of Locating Ferry In Integrated Linear Regression And Sequential NMS Scoring

With the rapid development of Computer Vision,the hotspot research of object detection has a very wide range of applications in the field of video intelligent monitoring.In traffic safety field,the accurate detection of the vehicle is the key of making the video intelligent monitoring in high level.So the ferry detection is the most important factor in Inland River Navigation Water Intelligent Video Monitoring.And obviously,the study of moving ferry has a very important significance.However,in the actual ferry detection process,the shape and scale of the vessel will continue to change,the dynamic background environment is also more complex,so the current ferry detection method are limited by the speed or accuracy.Most of the ferry Real-time accurate detection can't achieve the true sense of the level of intelligent video surveillance.Therefore,this paper proposes a method based on linear regression to realize the accurate detection of ferry and specific research work is as follows:First of all,the method of the Aggregate Channel Feature(ACF)is used to locate the ferry by using the aggregate channel feature which has the advantages of feature extraction,and we can obtained the ferry target detection box,and then training the position bias between the pretreatment detection frame and the ground-truth,so we can get the optimized model.Using the model to further adjust the ferry detection box' location and we can achieve precise positioning;Secondly,During the actual ferry inspection process,the images between the adjacent frames of the ferry video are very similar.Therefore,it is proposed to make full use of the related information between the front and rear frames of the ferry video to set the confidence score of the ferry target aim to reducing the error detection problem,and improving the accuracy of ferry detection.The experimental results show that the effects of ferry detection has been improved with the use of linear regression and sequence NMS.Compared with the traditional method of Aggregate Channel Features,the proposed method only lost 6 frames per second,and the accuracy rate was increased by 8 percentage points.Also compared with the convolution neural network which has the result of 20FPS,the linear regression of this paper reached the speed of 32FPS.But for mAP performance,the linear regression is only 3 percentage point less than convolutional neural network of 83.7.Therefore,the detection effect of this method improves the accuracy of ferry detection on the basis of ensuring real-time performance.