Research On Pedestrian Detection Method In Complex Weather Based On Deep Learning

Pedestrian detection technology is an important research branch in the field of computer vision,which is widely used in vehicle assisted driving,intelligent monitoring,intelligent traffic and many other aspects.In view of the problems such as low detection accuracy and large model encountered in pedestrian detection by traditional machine learning methods,deep learning was applied in the field of pedestrian detection in this paper.The Yolov5 algorithm of end-to-end detector was adopted as the basic method of pedestrian detection.A network structure of "high precision and lightweight in complex weather" was designed and its effectiveness was verified through experiments.The specific work of this paper is as follows:(1)In view of the problems of multiple parameters and large model of pedestrian detection algorithm,two lightweight deep convolutional neural networks based on Yolov5 are constructed in this paper.First,based on the idea of lightweight module design,linear transformation is used instead of convolution to generate redundant feature graph,and the calculation amount is reduced by designing different modules of feature extraction network and feature fusion network.Second,based on the idea of model pruning,the method of removing unimportant connections was adopted,and55% pruning was carried out on the standardized layer of the model batch to improve the major problems of the model.The experimental results show that the two lightweight networks proposed in this paper have fewer model parameters and computation,and their performance is better than other detection methods and lightweight methods.(2)In order to solve the problem of low accuracy of pedestrian detection,a pedestrian detection algorithm integrating attention mechanism is proposed.Firstly,the algorithm can enrich the local features of different data sets by the difference fusion method of different data sets.Then the coordinate attention module is used to capture channel information and global location information to enhance the feature extraction ability of the network.Finally,the vector Angle cost is added to the regression loss function to improve the training speed and reasoning accuracy of the model.By analyzing the performance of the proposed algorithm on the improved Caltech pedestrian data set,it is verified that the proposed algorithm can effectively improve the overall performance of the pedestrian detection network.(3)In order to further improve the accuracy of pedestrian detection in complex weather and make the algorithm applicable to sunny,rainy,snowy and foggy weather under natural conditions,this paper proposes a pedestrian detection network in complex weather.The network first uses the lightweight classification algorithm Mobilnetv3 to classify different weather conditions,and then input to the corresponding image processing module to complete the image to rain,snow,fog processing,and finally input the restored clear picture into the pedestrian detection network,constitute the pedestrian detection system under complex weather.Through experimental verification,the classification algorithm in this paper achieves the classification accuracy of 98.5%,and the pedestrian detection accuracy increases by16.2% compared with complex weather.