Research On Pedestrian Detection And Pose Recognition For Unmanned Driving In The Park

As an important research direction of driverless,park driverless vehicles have a wide application prospect in unmanned sales,logistics distribution,security patrol,passenger carrying,and other fields.As an important algorithm of park unmanned vehicle assistant driving system,accurate pedestrian detection and pose recognition results ont only can optimize the decision-making of unmanned vehicle but also can protect pedestrian safety.The accuracy of existing pedestrian detection and pose recognition algorithms based on deep learning have been significantly improved.However,this method generally have the disadvantages of large network model,large amount of calculation,long detection time and complex model design.Therefore,in order to reduce the network model while ensuring detection accuracy,this thesis studies and improves the pedestrian detection and pose recognition algorithm based on lightweight model.The main contents of this paper are as follows:To overcome the problems such as high missing rate and low detection accuracy of TinyYOLOv3 algorithm,an improved algorithm is proposed.The depth separable convolution is used to replace the traditional convolution,and the features of different levels are fused to increase the number of network layers.Experimental results show that this design method reduces the amount of network parameters while ensuring recognition accuracy.In the improvement of the loss function,the CIoU loss is used to replace the original bounding box coordinate to predict loss.The channel attention module is integrated into the feature extraction network.The experimental results show that CIoU and the channel attention module are helpful to improve the accuracy to a certain extent and the recognition accuracy.The data enhancement method is used to avoid over-fitting training,and super parameter optimization,dynamic learning rate setting,prior frame clustering,and other methods are used to accelerate network convergence.This study uses INRIA and VOC fusion pedestrian dataset for verification experiments.The experimental results show that the missing rate of the improved algorithm is reduced by 23%,the size of the network model is reduced by 5.6MB and the value of m AP is improved by 12.92% compared with TinyYOLOv3.In terms of model size and accuracy of pedestrian detection,the improved algorithm achieved better results.We proposed a pedestrian detection algorithm based on multi-scale fusion feature.The information of different scales is fused by redesigned the main network with multi branch convolution module.In order to improve the nonlinear ability of the network,the Mish activation function is used for network design.We improved the FPN structure by adding a cross connection in the output part of different scales.The prediction output of three scales is used to strengthen the network’s ability to detect small-scale pedestrians.Compared with TinyYOLOv3 algorithm,although the detection time of proposed algorithm is slightly increased,the network model is only 17.7M,and the detection accuracy is significantly improved.Pedestrian pose recognition based on human skeleton image is studied.Traditional convolution neural network is prone to over fitting when processing human skeleton data,in order to solve this problem and realize pose recognition based on human skeleton image,a lightweight pedestrian pose recognition algorithm based on improved Ghost Net network is proposed.Firstly,the input size of the network is modified according to the characteristics of the human skeleton image,and the 5×5 convolution kernel is used to enhance the receptive field.Secondly,suppress overfitting by reducing the number of feature layers and network layers.Finally,the designed network is compared with some classic networks.The experimental results show that the proposed algorithm achieves better performance with a small amount of network parameters.