Software Development Of Intelligent Camera System Based On Hi3559

With the development of artificial intelligence and big data,intelligent Internet of Things products have gained widespread popularity and market recognition,also facing the challenges of massive data transmission,computing and storage.Compared with traditional webcams,smart cameras can perform intelligent analysis tasks at the source of data collection,assist rapid decision-making,reduce network bandwidth occupation,and has important engineering application value in building decentralized,real-time intelligent video surveillance system.Accordingly,this thesis develops a smart camera system software based on Hi3559 chip for real-time intelligent video surveillance,uses Hi3559's AI coprocessor to develop real-time intelligent analysis algorithms and adopts a modular approach to divide the system into three parts: system business module,video service module and intelligent analysis module.The system business module provides the equipment operation and maintenance interface through the web page,and uses a proprietary session and upgrade protocol to realize interactive functions such as video preview,parameter configuration,and application upgrade.The video service module implements real-time video capture and encoding,as well as transmission based on mainstream RTSP and HTTP-FLV streaming media protocols,and provides personalized stream configuration functions.In the intelligent analysis module,the thesis designs an extensible intelligent analysis software framework with Hi3559 neural network inference engine as the processing core.Based on this framework,developers can quickly and efficiently deploy intelligent analysis tasks,which perform real-time analysis and annotation on the collected video frames in a cooperative manner of software and hardware.Face recognition and pedestrian detection are deployed as cases.The former uses Retina Face,Mobile Face Net network and NMS algorithm,affine transformation,cosine similarity and other technologies,the latter uses Faster R-CNN two-stage network and region proposals' bounding box regression technique.Test results show that the system is qualified for the acquisition and transmission of 1080P/30 fps real-time video stream and provides a self-contained HD camera configuration function.The real-time performance and accuracy of face recognition and pedestrian detection deployed in the intelligent analysis software framework can meet the application requirements of certain scenes.