Research on Intelligent Target Detection and Coder-decoder Technology Based on Embedded Platform
Xiaodong Zhao, Xunying Zhang, Xuemei Cheng, Fayang Chen, Zuofeng Zhou, Tao Xu · 2019
In order to meet the embedded application requirements of machine learning algorithm, the intelligent target detection and recognition algorithm based on convolutional neural network and corresponding optimal process are studied. Detailed network structure analysis and network performance analysis are carried out. Based on GPU embedded platform, TensorRT technology is used to accelerate the embedded application of intelligent target detection and recognition algorithm, including fp16 and int8 inference modes. Satisfactory verification results are achieved on embedded platform. In addition, an integrated system of real-time machine learning and H.265 encoding and decoding technology is realized. Firstly, the compressed image data sent by the camera is received by embedded platform and decoded in real time in H.265 format. Then the real-time intelligent target detection and recognition algorithm basing on TensorRT technology is done for RGB data obtained by hardware decoding process. Finally, the data is compressed in H.265 format, and subsequently storage and data transmission are carried out. The experimental results show that TensorRT technology can improve the inference speed of neural network in embedded platform. The network structure optimized by TensorRT technology can achieve three times the speed increase, with limited accuracy loss. Hardware coding and decoding of H.265 can also cause corresponding delay to program inevitably.