Research On Target Detection Application Based On RK3399Pro And YOLO

In recent years,the computational requirements of target detection algorithms represented by the deep neural network have increased exponentially.In a wide range of fields,demands for deployment to embedded platforms also increase steadily.However,the related algorithms are difficult to be deployed in real-time embedded platforms with limited performance.Besides,the commonly used AI platforms for science are mastered by foreign manufacturers.It faces the phenomenon of "stuck neck" in core technologies and key areas.There are information hidden dangers in applied research.Therefore,the research on the deployment technology of target detection algorithms for domestic intelligent platforms should be endowed with important application value.The work of this paper is mainly carried out from the following aspects:1.For the demand of developing an intelligence processing platform based on domestic chips,the overall architecture and implementation scheme is designed.Video capture,display,and forward calculation processing module also are planned.In this paper,Rockchip’s RK3399 Pro is selected as the prototype of the embedded intelligent processing platform.The hardware construction is completed,an integrated development environment is built,and an embedded system with stable and efficient data processing capability is constructed.2.The target detection algorithm is deployed and optimized based on the developed intelligent processing platform.For the constraints of the embedded platform with limited computing and storage resources,the algorithm is adapted to improve.And various strategies are adopted to reduce the running time.When the image resolution is 416 pixels × 416 pixels,Tiny YOLOv4’s speed of theoretical reasoning reaches 80 FPS,by modifying the number of convolution channels and replacing the activation function,etc.YOLOv5 s breaks through the limitations of the RKNN framework on the neural network layer.Its theoretical inference speed reaches50 FPS,by using the improvement of feature extraction and model pruning,etc.Both run more than 2.2 times faster.3.For the application of intelligent technology in engineering,a real-time target detection system is constructed based on the intelligent processing platform and optimized algorithm.Besides,the user interface is designed,and the performance of this system is verified by UAV detection experiment.To sum up,for the demands of domestic intelligent processing system,this paper has carried out the development of intelligent processing platform,target detection algorithms,and UAV detection system.When the image resolution is 416 pixels ×416 pixels,the target detection processing capacity reaches 60 FPS,which meets the real-time application requirements.