Human Frontal Instance Segmentation Based On Nvidia Jetson Platform

Affected by the COVID-19,it is very important to deploy a mobile temperature measurement system in public places.Aiming at the problem that the existing infrared temperature measurement system cannot accurately select the temperature measurement area of the human forehead,this paper proposes a method for segmentation of the human forehead instance based on the Nvidia Jetson platform.Instance segmentation algorithm,as an image processing algorithm that combines target detection and semantic segmentation,can accurately segment and frame different targets in the detection area.Improved instance segmentation algorithm solves the problem of adhesion and loss of part of the human body temperature measurement area segmentation;for the improved The instance segmentation algorithm occupies too much storage space.A model compression algorithm based on the vector quantization(VQ)method is proposed,and the search algorithm of the permutation matrix is improved to further compress the model size to facilitate deployment to the embedded platform;finally,the compression is based on the Tensor RT framework The latter model is parsed into Tensor blocks to achieve the purpose of improving the calculation performance of the embedded instance segmentation algorithm.The main research contents of this article are:(1)Aiming at the problem of adhesion and incompleteness of segmented targets,an improved algorithm FFPN(Fused Feature Pyramid Network)based on FPN(Feature Pyramid Network)in Mask R-CNN is proposed.In-depth study of the realization principle of the case segmentation algorithm,analysis of the realization principle of FPN network extraction features.Refer to the neural network of skip connection structure,design the semantic information transmission channel from the upper layer feature map to the lower layer feature map.The features of each channel are merged to enhance the semantic information of the underlying feature map,which improves the accuracy of the network by about 8%.(2)Aiming at the problem that the volume of the improved instance segmentation model occupies too much storage space.Research on model compression algorithm based on feature matrix quantization.Improve the vector quantization compression algorithm applied to the convolutional neural network,introduce the rate distortion theory to search for the best permutation matrix to improve the compression ratio of the model,and introduce the annealing k-means algorithm to adjust the neural network to reduce the accuracy loss of the quantized model.(3)Implement algorithm deployment based on Jetson NANO.Research on the hardware and software computing acceleration principles of Jetson NANO embedded computing platform,and adjust the hardware and software settings of the hardware platform to make it suitable for deep learning model acceleration computing.Research the Tensor RT acceleration framework conversion model,and output the deserialized Tensor model that can be inferred in the Tensor Engine framework.Experiments show that the efficiency ratio of this scheme is much higher than that of the graphics workstation,and the instance segmentation effect reaches the expected.