Research On Fisheye Image Target Detection Based On Deep Learning

Fisheye lens is a kind of super wide-angle lens.Compared with the general image,the fisheye image has the characteristics of large viewing angle,which has more information.Therefore,it can use less acquisition equipment with fisheye lens to obtain a larger range of scenery,which can simplify the image acquisition process and reduce the waste of hardware resources.Because of the above characteristics,fisheye lens has been widely applied in video conference,security monitoring,intelligent transportation,astronomical observation,medical detection and other fields.However,fisheye lens has serious distortion,which makes,and reduces the application of fisheye lens.Based on the fisheye camera,firstly,this thesis investigates the calibration method based on the calibration chessboard,which is to calibrate the corner points by collecting chessboard of different angles,and obtain the parameters of the fisheye camera.This method is flexible and easy to achieve,does not rely on complex calibrators and has high calibration accuracy.Secondly,due to the shooting angle and light condition,the fisheye image will be distorted,rotated and stained.The existing correction methods cannot obtain the satisfied image correction.Therefore,this thesis proposes a method combining the latest temporary interpolation with Transform Invariant Low-rank Texture(TILT)to correct the fisheye image.The proposed method can correct the correct the fisheye image and the target angle of the image with the characteristics of high correction accuracy and simple implementation,which has good practical value.Thirdly,based on the YOLOv3 detection algorithm,the target detection is carried out,and the experimental results show that the recognition accuracy of corrected fisheye image is higher than that of regular fisheye image.In order to further improve the detection ability of the algorithm for the distorted fisheye image,this thesis proposes an improved YOLOv3 algorithm to improve the feature extraction ability.The improved algorithm has a higher resolution for the same kind of distorted target features and can depict more accurate position information in comparison to YOLOv3 algorithm.The experimental results show that the improved algorithm has a significant improvement of recognition accuracy,and can also meet the needs of practical applications.Finally,a real-time visual detection interface is designed by integrating camera calibration,image correction and target recognition.The functions of the whole system are integrated on the interface to verify its feasibility.The system realizes the online real-time target detection based on fisheye vision,and the operation is convenient.