Research On Uav Reconnaissance And Processing System Based On Tegra

Real-time performance is a very important parameter in uav reconnaissance system.This uav reconnaissance and processing system based on Tegra will be optimized for the real-time performance of the system to achieve the effect of ultra-low delay.The details are as follows:（1）Optimize the deployment reasoning of the algorithm.When the trained model is deployed,the equipment at the deployment end is far less equipped than the training equipment,which leads to problems such as slow reasoning speed,high delay time and poor real-time performance.To solve these problems,the system adopts model precision transformation and zero copy optimization.Firstly,the accuracy of FP32 data during training is reduced to that of FP16 data,thus reducing the amount of data calculation and speeding up the calculation speed.Secondly,zero copy is adopted to cancel the process of data copying between CPU memory and GPU memory,so that all the time is spent on GPU computation to reduce the inference time.（2）Design the system HEVC coding.As HEVC coding increases the compression rate,the complexity of the algorithm also increases greatly,which makes it difficult to carry out real-time HEVC compression.To solve this problem,this design improves the compression speed by means of hardware acceleration,so as to realize the real-time coding effect of HEVC.At the same time,in order to improve the versatility of HEVC encoders,video format conversion is accelerated before coding,so that HEVC encoders support a variety of different formats of video encoding.In order to ensure that the video format conversion does not affect the real-time coding of HEVC,GPU parallel acceleration is also carried out to realize the real-time coding module of HEVC that supports multiple formats.（3）The UAV reconnaissance and processing system based on Tegra is realized.The system includes video capture,TensorRT deployment reasoning,HEVC compression coding,UDP network transmission and upper computer reception.The optimization acceleration of TensorRT deployment reasoning and the optimization design of HEVC coding are realized.Finally,the UAV reconnaissance and processing system based on Tegra is realized with an ultra-low delay of about 250ms.