Research On Object Detection Based On 4D Light Field Computational Imaging Technology

With the rapid development of artificial intelligence,image-based object detection technology has been widely used in the fields of unmanned driving,unmanned shops,and intelligent transportation.However,the accuracy of the traditional image-based object detection method is limited by the quality of the image.Once 2D image appears out of focus,blurring,and occlusion,it will directly affect the accuracy of the object detection,and the traditional 2D image cannot detect the depth of the object.As an important branch of computational imaging,Light Field has caused a wave of research in computational photography in recent years.Different from the traditional imaging,the Light Field can record both position and angle information of the light at the same time.Reconstructed by Light Field algorithm,Light Field can realize digital refocusing,digital change the depth of field,viewpoint conversion and depth estimation,etc.Object detection based on Light Field can achieve object depth information while improving object detection accuracy.However,there are many shortcomings in Light Field such as low resolution,aliasing,and slow calculation speed,which seriously hinder the rapid promotion and application of light field.Therefore,in order to improve the accuracy and speed of Light Field,the focus of Light Field and synthetic aperture imaging are studied.First of all,this paper proposes an de-artifact algorithm based on the sliding window and adaptive filtering to improve the imaging quality of focused Light Field.Secondly,for synthetic aperture Light Field,in order to improve the deficiencies of affine reprojection in traditional disparity calibration methods,a layered calibration method based on perspective reprojection is proposed to improve the calibration accuracy.In addition,a synthetic aperture algorithm based on layered calibration method is proposed and accelerated by GPU.In the depth estimation of light field,this paper improves the shortcomings of traditional defocus depth estimation in slow speed and high noise,and proposes a fast local depth estimation scheme based on defocus response.In terms of experiments,this paper designed and built a camera array system to collect real four-dimensional Light Field data,and obtained a focused stack using an improved synthetic aperture algorithm.On this basis,the YOLO and face recognition engine were used to detect the focus stack,and a non-occluded and occluded object detection experiment is performed for the synthetic aperture Light Field.The experimental results prove that Light Field based object detection not only can obtain object depth information,it can even achieve "perspective" of sparse obstructions to detect object.