Research On Vanishing Point Detection And Distortion Correction Method For Vehicle Recognition

With the construction and development of smart cities,a large number of surveillance cameras are widely used in scenes such as illegal capture and vehicle tracking.The main task is to realize vehicle identification in urban traffic,and realize the detection and statistics of traffic flow and vehicle inspection through detection and identification of vehicles,which plays an important role in intelligent transportation.In vehicle recognition,the effective calibration of the traffic bayonet camera is the key to realize the above-mentioned applications.The subsequent observation and measurement tasks require the internal and external parameters of the camera calibration.By correcting the distortion of the camera and solving the internal and external parameters,the accuracy of subsequent tasks can be effectively improved,thereby improving the performance of vehicle recognition.The automatic detection of the vanishing point of an image is the main prerequisite for fully automatic camera calibration,and the internal parameters of the camera can be solved by the vanishing point of the image.At present,in camera calibration,the commonly used vanishing point detection methods are mainly divided into texture-based methods and edgebased methods.The edge-based methods are fast,but rely on strong edges in the image to cause low accuracy.The methods related to distortion correction are mainly divided into traditional correction method,active vision calibration method and self-calibration method.The selfcalibration method uses the relationship between the corresponding points of multiple views for calibration.This method is flexible and convenient,but it is computationally complex and has poor robustness.In response to the above analysis,this article has specifically studied the following contents on the basis of fully studying the existing work:First,for the problem of vanishing point detection at urban traffic checkpoints,this paper proposes a vanishing point detection method based on constraint classification.In the detection of straight lines,according to the relationship between the length of the straight bars and the pixels in the image and the chaotic transformation law of straight line angles,a short straight line filtering method is proposed.By selecting an appropriate short straight line filtering threshold,the messy short straight line data is filtered out.Then combined with the image characteristics and the similar change law of the straight line angle,the hierarchical clustering method is used to screen and classify the straight lines,and the vanishing point is obtained by quantifying the diamond space and voting to select the maximum value.The experimental results show that compared with the classic method,this method not only greatly reduces the computational complexity of the image vanishing point,but also improves the accuracy of the image vanishing point detection to a greater extent.Second,aiming at the problem of camera radial distortion correction,based on the principle of linear projective invariance,a self-correcting method of radial distortion based on edge linear features is proposed.First,the straight line subsets with the same edge characteristics are solved,and then the relationship model between the straight line and the radial distortion coefficient is established according to the collinear condition of the straight line subset,and the first-order distortion coefficient is obtained by iterative optimization using the quasi-Newton method.By establishing a reasonable constraint function,the distortion coefficient and distortion center are optimized.Finally,the radial distortion is corrected and restored,and the vanishing point location is obtained using the vanishing point detection method in the previous chapter,and then the internal parameters of the camera are solved,so as to realize the self-correction of the single image camera.Experimental results show that,compared with the traditional correction method,this method is based on a single image,does not require specific calibration objects and manual intervention,and is simple to operate and has higher accuracy.Third,in order to verify the effectiveness of vanishing point detection and distortion correction in vehicle recognition in camera calibration,this paper uses vanishing point detection and distortion correction to process different data sets(KITTI and BITVehicle),and use the SSD-based vehicle recognition model to verify.Experiments show that the performance of the vehicle recognition model has a certain improvement in the test set after the above-mentioned processing compared with the image without any processing.