Research On Video Conversion And Recognition Technology Based On Embedded System

Video surveillance is an important application technology field for social security,security,security and people's livelihood.With the development of image processing technology,digital and intelligent video surveillance system has become mature and widely used.Video structured processing technology which intelligently analyzes original video and extracts key information has achieved good practical results in the field of video monitoring.At present,most of the mainstream video structured processing schemes use highperformance GPU,but the cost of platform installation is high,especially for the equipment update and performance improvement of the old video monitoring system,which requires a lot of capital investment.In view of this,a low-cost embedded video structured processing system without replacing the original video monitoring device is proposed,that is,the video streams monitored are structured on the embedded chip by connecting the video interface with the existing monitoring device.In this paper,the algorithm and technology of video conversion and video recognition are discussed.For video conversion,this paper studies the conversion between VGA and HDMI interfaces used in monitoring and display devices to the embedded video input protocol BT.1120.Firstly,the video conversion from VGA interface to BT.1120 protocol is studied and implemented from two aspects: color space and time sequence synchronization of VGA interface.Then,according to the structure of HDMI interface signal,the conversion from HDMI interface signal to BT.1120 transmission protocol is studied from the perspective of HDMI data cycle.According to the resolution and functional requirements,a video conversion unit with the VGA and HDMI interface conversion core of TVP7002 and ADV7610 is designed.For video recognition,the imaging principles of traditional target detection algorithm and convolution-based neural network target detection algorithm are studied,and the shortcomings of traditional algorithm are analyzed.Based on the comparison of algorithms of main convolution neural network and the limitation of computing resources of embedded platform,YOLOv3-Tiny algorithm with both accuracy and real-time is selected as the basic network.For the characteristics of targets in VIDF data set,an initialization scheme based on k-means clustering algorithm is proposed,and the number and size of anchors suitable for targets in this dataset are calculated.To overcome the drawbacks of YOLOv3-Tiny algorithm in small target feature extraction,the feature extraction capability of the algorithm is improved by increasing the number of convolution layers,and the NIN convolution layer is inserted to reduce the computational complexity.The experimental results show that the improved algorithm can significantly improve the accuracy of YOLOv3-Tiny algorithm on the premise of satisfying the real-time target detection.The video conversion and video recognition are implemented in the Hi3519 A embedded platform.Verify the feasibility of video conversion unit implementation process through Hi3519 A video input port printing information verification and laboratory equipment inspection method.The Darknet framework of yorov3 tiny algorithm is transformed into the Caffe framework and transplanted to the hi3519 a embedded platform.The test results verify the improvement of the detection effect of the improved YOLOv3-Tiny algorithm.