Pedestrian Detection And Tracking Of Fisheye Images Based On Multi-feature Fusion

Pedestrian detection and tracking has always been a popular field of computer vision research,in video surveillance,military technology,intelligent traffic management and so has a very wide range of applications.At present,most of the computer vision system used in ordinary images of the collected images of the field of view is generally 40 degrees to 50 degrees,If only a camera is difficult to obtain all the surrounding information,making the scope of the observation is limited.Select the fisheye lens can achieve a single point of a wide range of monitoring,for pedestrian detection and tracking will be able to obtain a larger area of visual scene.But the fisheye lens image will be accompanied by a greater degree of distortion,in the pedestrian detection and tracking time will encounter some difficulties.In this paper,the application of pedestrian target detection and pedestrian tracking algorithm in fisheye lens is studied,and the characteristics of fisheye image are improved and the correction of fisheye image is studied.The main results are as follows:(1)For the detection of moving pedestrians in fisheye images,this paper first analyzes some classical moving target detection algorithms including background difference method,optical flow method and interframe difference method.For the three methods of the principle of research and experimental analysis,according to the actual situation finally put forward a hybrid Gaussian model based on the background difference method of foreground extraction method.First,the background of the fisheye image is updated in real time using the mixed Gaussian model,and then the moving target of the foreground is obtained by aligning each frame image with the background.According to the different size of the moving target caused by the lens distortion in the fisheye lens,the improved morphological processing of the fisheye image is based on the geometrical characteristics of the fisheye image distortion,and the foreground target of all the fisheye images is obtained.For the extracted foreground targets,the simple SVM classification is performed with the HOG feature to determine whether or not pedestrians are,and finally the complete pedestrian of the fisheye image is obtained.Experiments show that the improved algorithm has good real-time and robustness for the extraction of moving pedestrians in fisheye images,and it is fast and has high hardware requirements and high feasibility.(2)For the tracking of pedestrians in fisheye images,this paper first studies several commonly used tracking algorithms,and on this basis,aiming at the multi-target tracking in fisheye images,a method based on Kalman filter Pedestrian tracking algorithm combined with color texture and other features of the multi-target tracking framework.After a specific experimental verification,we propose this multi-target tracking framework in the fish-eye video can effectively track multiple sports different goals,with a relatively high accuracy of pedestrian tracking.(3)The availability of the effective area of the fisheye image and the correction of the image are two important aspects of the realization of the fisheye image processing algorithm.For the extraction of the circular effective region,this paper analyzes the advantages and disadvantages of several commonly used effective region extraction algorithms,such as the progressive row-by-row scanning algorithm,the area statistic algorithm and the regional growth algorithm,according to the characteristics of the video and the real-Requirements,the use of progressive row by row scanning algorithm to extract the effective area.For the moving object detected in the fisheye image,it may be necessary to extract the observation,but because of the distortion of the fish-eye image,we need to correct it.In this paper,a simple and practical method is proposed,which uses the characteristics of the fisheye image itself to find the spherical radius and the optical center.Then,the fish image is corrected to the normal image by the simple mathematical model.Through the experimental verification of the fisheye video,we propose a method to detect and track pedestrians in fish-eye images accurately and efficiently.