The Research Of Abnormal Condition Warning For Driving Assistance Based On Machine Vision

With the development of artificial intelligence,more and more new technologies have entered our life.The driving assistance system of automobile is one of them.The research of abnormal condition warning for driving assistance based on machine vision can not only be an effective way to save drivers live,but also reduce traffic accidents.And it is also the basis of unmanned vehicle technology.Comparing with those detection technology based on radar and LiDAR,machine vision has its own advantages,such as more accurate in detection,more useful for human visual or machine perception,lower cost,strong applicability etc.Some innovations are proposed in the algorithm structure and implementation after summarizes the previous research on object detection,lane detection and monocular vision ranging.Then this paper presents an early warning system with high accuracy,real-time and robustness.This system mainly includes two parts,object detection and tracking part and early warning part.In object detection and tracking part,I select the Faster R-CNN for object detection and solve the problem of incorrect detection result by using cascade classifier.The KCF tracking algorithm is combined with the Faster R-CNN to improve the tracking speed.In the early warning part,the use of lane line detection based on the ridgeness not only increases the detection success rate,but also can detect lane line in only one side.In the end,the whole road can be modeled and the warning is given more accurately by using the single visual range.The innovations of this paper is as follows.Firstly,the network of Faster R-CNN was optimized in the structure of the last convolution layer and pooling layer.Secondly,the KCF is extended by multi-template scale-adaptive,which improves the tracking effect.Additionally,different from the previous lane line detection algorithm based on single feature or multi-feature.This paper presents a novel detection method based on the ridgeness.An effective noise filtering mechanism will next remove noise pixels to a large extent.A modified version of sequential RANS AC is then adopted in a model fitting procedure to ensure each lane line in the image is captured correctly.Finally,the new ranging formula is derived by the 3d ranging model.The error detection rate is reduced by eliminating the calculation of the internal parameters of the camera.