Design And Implementation Of Intelligent Camera Based On Hi3519A

With the development of embedded artificial intelligence chip and deep learning,the application of intelligent camera is becoming more and more extensive.The intelligent camera can directly complete the intelligent processing of image data locally,saving a lot of processing time and labor costs.Especially in the field of aerospace,the intelligent cameras on aerospace vehicles and aerial photography drones need to complete the monitoring of the working status of various parts,the identification and tracking of external flying objects,and the detection of ground targets,but the aircrafts have very strict requirements on the installation space of various equipment.Therefore,the research of miniaturized and intelligent camera has high practical value.In this context,an intelligent camera based on Hi3519 A is designed.According to the characteristics of intelligent camera miniaturization,the hardware circuit is divided into acquisition module,power module,main control module and Ethernet module,and the FPC is used to connect these modules.The IMX421 image sensor of the acquisition module converts the collected original video data through AD and transmits it to Hi3519 A image processing main control module by SLVS-EC interface.After the intelligent processing is completed,the data is transmitted to external devices through gigabit Ethernet module.The power module ensures the power supply of each module.According to the characteristics of camera intelligence,the software design is divided into video transmission module and intelligent processing module.The video transmission module completes the basic video playback function.Video is transmitted to the client by the video input module,H.264 encoding and RTSP protocol.The intelligent module completes the object detection of the image.The hardware acceleration unit NNIE of Hi3519 A is used to deploy the intelligent algorithm,and the Faster R-CNN is used to realize the object detection of aircraft and ground targets.Faster R-CNN was trained in the weak aircraft dataset and RSOD dataset,and m AP was91.35% and 91.10%,respectively.It has been verified that the size of the smart camera after assembly is 50×50×30mm,and the real-time video playback function of 1080P@30fps has been realized.The trained network model is transplanted into the camera,and the image is detected by NNIE.Although the detection result is slightly lower than that of PC,the intelligent camera has good accuracy.