Software Development Of Panorama Stitching And Pedestrian Detection Svstem Based On Hi3559

Video surveillance systems have been increasingly used in our business and personal lives.However,the independent and pose-fixed cameras in traditional systems have limited camera fields of view.They leave potential security risks on the blind spots.Besides,the current system requires human security personnel to monitor and analyze the videos.The stability and reliability of those analyzers decrease when pedestrian occurs densely.Video stitching and object detection are two direct solutions to the problem.Video stitching explores the overlapping areas of multiple videos and produces panoramic videos with a wider field of view,while object detection recognizes moving objects in real-time.In this thesis,this article embedded the two methods into Hi3559,a professional 8K ultra-HD mobile camera System on a Chip(SoC)and developed three modules in the systems:streaming media processing(SMP)module,video stitching(VS)module,and pedestrian detection(PD)module.(1)SMP module preprocesses the video data from three channels.It also encodes the panoramic video data after the VS module and hosts it for browser access.(2)VS module stitches the video using hardware acceleration.It extracts the feature points of the video frame based on the Speeded Up Robust Features(SURF)method,matches the feature points by top-2 nearest neighbors,distills the feature points from RANdom SAmple Consensus(RANSAC),and stitches the three-channel video using the optimal seamline theory.(3)PD module detects foreground moving targets based on Gaussian Mixture Model(GMM).It uses the IVE module of the development board to implement GMM modeling,foreground detection,post-processing,connected area calibration,and hardware acceleration.The test results proved that the system can complete real-time stitching and pedestrian detection on three-channel videos while maintaining a good performance.The system gives a high visual quality of the stitching video.The average stitching time per frame is 33.8 ms and 34.4 ms for indoor and outdoor scenes,respectively.Our system also achieved a decent accuracy on pedestrian detection.The average processing time is 31.8 ms per frame,which meets the real-time requirements of the surveillance system.The system carries promising values in engineering applications.