Research On Gesture Recognition For Human-Vehicle Interaction Based On Lightweight Convolutional Neural Network

Human-computer interaction based on gesture recognition has the characteristics of natural,intuitive and efficient.When applied to intelligent driving,the in-vehicle devices are controlled by gestures,which reduces the visual attention and the time when the driver takes off the steering wheel for human-vehicle interaction,to improve driving safety.Under vehicle conditions,memory and computing power are limited,and the collected gesture images are affected by complex backgrounds,lighting changes,shadows,and vehicle bumps.Therefore,this paper proposes a human-vehicle interaction gesture recognition algorithm based on lightweight Convolutional Neural Network,aiming to achieve a high level of gesture recognition accuracy and model lightweight under vehicle conditions.The main research contents are as follows:1.The gesture recognition algorithm based on traditional vision and convolutional neural network is analyzed and researched.According to the characteristics of human-vehicle interaction gesture recognition,the overall framework of the human-vehicle interaction gesture recognition algorithm is designed,which includes gesture localization algorithm and gesture classification algorithm.2.The lightweight network YOLOv3-tiny is selected as the gesture localization platform,the localization accuracy improvement of this network is studied,and a gesture positioning algorithm based on YOLOv3-tiny is proposed.The network accuracy improvement method includes: stacking 3×3 convolutional layers in the backbone to increase the network receptive field;introducing a Cross Stage Partial Network in the backbone to improve the network learning ability;integrating the Spatial Pyramid Pooling module in the network neck to reduce the influence of the variety of gesture sizes and positions on the accuracy;K-means clustering and linear transformation are applied to adjust the size of the anchor box to improve the matching degree of the anchor box and gesture.3.The lightweight network ShuffleNet V2 is selected as the gesture classification platform.In order to improve the lightweight degree of the model under the premise of maintaining high accuracy,the structure of the network is adjusted,and a gesture classification algorithm based on Shuffle Net V2 is proposed.Network structure adjustment methods include: remove the redundant 1×1 convolution layer in the network unit,reduce the number of superpositions of the basic unit,thereby reducing the network redundancy;expand the convolution kernel size of the depth separable convolution to improve the network receptive field;insert 3×3convolutional layer into the network to improve the feature extraction ability of the network.4.Conduct experimental analysis on gesture localization algorithm,gesture classification algorithm and human-vehicle interaction gesture recognition algorithm.The experimental results show that: the improved YOLOV3-tiny increases the average accuracy by 6.32% with a small number of parameters and inference time added;the improved Shuffle Net V2 maintains a high accuracy rate,and the inference time and parameters are reduced by 30% compared to the original network;the accuracy rate of the human-vehicle interaction gesture recognition algorithm on the NUS-II gesture dataset reaches 99.46%,the detection time is only 15 ms,and the number of parameters is only 10.8M.Compared with other current algorithms,it not only has a higher accuracy rate,but also achieves a higher real-time and lightweight level.