Method And Implementation Of Moving Object Detection Under Motion Imaging Platform Based On Image Registration

Moving object detection under a moving platform refers to the process of detecting moving objects in the image sequence acquired by the moving imaging platform and separating the moving objects from the motion scene.It is a basic procedure in analysis and understanding of videos for computers,which is used in many applications,such as surveillance,intelligent assisted driving,robot environment perception,photoelectric investigation,precision guidance,and so on.Detecting moving objects accurately and effectively is the foundation of the above-mentioned high-level visual tasks,however,it faces many challenges,including 1)It is difficult to extract and model the feature accurately under complex motion scenes and imaging conditions;2)It is difficult to accurately separate the object from the background when the camera moves,since the motion of the object and the background are coupled with each other;3)It is difficult to implement a real-time object detection system in conditions of limited resource constraints.To deal with these challenges,this thesis focuses on image registration based moving object detection method and system,improves image registration accuracy and object detection accuracy with the accurate acquisition of background and object motion information,the prediction and estimation of image registration accuracy,the co-fusion method of image registration and object detection.Moreover,a high time-efficient objection detection system is implemented based on the multi-level,multi-scale parallel/pipeline heterogeneous parallelism architecture,which breaks through the platform resource constraints.Aiming at the decrease of feature matching performance and image registration performance caused by the main orientation estimation error of the feature point,the error source of existing methods for estimating the main orientation of the feature point is analyzed,and the key factors that affect the main orientation accuracy are studied.Based on these,a meaning-full orientation that can characterize the overall trend of the gradient is defined for feature points.And a new orientation estimation method is proposed,in which a rotation-invariant gradient is introduced to describe the spatial structural features in the feature region accurately.Experiments show that this method significantly improves the accuracy of the main orientation estimation,increases the proportion of feature points that could be matched correctly by 9%～30%,and effectively improves the number of point pairs matched correctly and registration accuracy in image registration.To deal with the insufficient motion decoupling and detection performance degradation caused by inaccurate image registration,an exploratory study on the prediction method of sequence image registration accuracy is launched.The performance of different feature extraction methods in sequence image registration are compared,and the characteristics of sequence images are analyzed,based on these,the key factors that affect the registration accuracy are studied,and a registration accuracy prediction method based on uniformity of inlier distribution is proposed.The proposed method models and analyzes the uniformity of the inlier distribution to predict and estimate the sequence images registration accuracy.Experiments on different feature extraction methods and different types of images show that the accuracy of predicting registration results is as high as 80% or more with our method,and most of the inaccurate registrations can be detected,providing valuable guidance for subsequent visual tasks.To improve the accuracy and real-time performance of moving object detection,the effect mechanism of image registration on object detection is analyzed,the mechanism of registration accuracy and detection accuracy decrease during object detection is studied,and a new moving object detection method is proposed,in which image registration is integrated with moving object detection.Unlike the traditional pipelined object detection framework,this method feeds the information of detection results back and uses the detection information of the previous frame image to improve the registration accuracy of the current frame image.Meanwhile,the accuracy of image registration is predicted and used to guide subsequent object detection,improving object detection accuracy.Experiments show that,with co-fusion of registration and detection information and feature extractor selection,this method significantly reduces false alarms and improves object detection accuracy and real-time performance.To further improve the real-time performance of the detection system,a real-time detection system for moving objects based on FPGA+DSP is proposed.On the basis of comprehensive analysis of calculation characteristics and computational requirements of each module of the detection algorithm,the algorithm structure is optimized and multi-scale and multi-level parallel/pipeline structure is designed to reduce the occupancy rate of FPGA and improve the real-time performance of the system.The system can detect moving objects in image sequence with size of 1280×720 under 60 frames per second,and the detection performance is only 0.02% different to the relevant pure software implementation,meeting the actual application requirements for power consumption,system size,processing speed and detection performance.