Research On Pedestrian Detection Algorithm For Railway Scene

Railway transportation is an important link of transportation in China and the main artery of national economy.With the progress of science technology and the development of economy,China’s railway network continues to expand,its stable operation not only maintains the national economic development,but also concerns the safety of people’s property.Railway foreign Invasion is one of the main threats to railway traffic safety,but the traditional security methods have many limitations.In recent years,deep learning has shown great development prospects.In this thesis,the object detection algorithm based on deep learning is applied to the Pedestrian detection task in the railway scene.SSD is taken as the basic detection algorithm,and in order to meet the practical application requirements,the adaptive improvement is made to it.The main work contents are as follows:(1)This thesis mainly aims at Pedestrian detection in railway scene,and improves the original algorithm from many aspects according to the characteristics of railway scene images.By integrating the semantic information of the deep feature map into the shallow feature map,the generated new feature map not only maintains more detailed information,but also has more abundant semantic information,which enhances the detection effect of the small-scale object.A multi-branching atrous convolution module was designed to obtain multi-scale information from the feature maps processed by the module,which enhanced the feature representation ability.Meanwhile,an attention mechanism is introduced into the module to optimize the feature extraction process.According to the characteristics of the object in the railway scene,the prior box responsible for classification and regression in the model was reset to give full play to the feature extraction ability of the improved algorithm.The experimental results show that the detection ability of the improved algorithm is enhanced.(2)In order to save the calculation and storage resources of railway monitoring equipment in the actual scene,this thesis applies the lightweight network to Pedestrian detection for railway scene.Efficient Net-B0 and Efficient Net-B3 were used to replace the SSD base network VGG16 respectively,creating Efficient Net B0-SSD and Efficient Net B3-SSD.The results showed that the replacement of the basic network significantly reduced the parameters and calculation of the model.Then,the feature fusion module and atrous convolution module designed in this thesis were embedded in the Efficient Net B0-SSD and Efficient Net B3-SSD to ensure that the algorithm still has a certain detection accuracy after reaching the goal of narrowing the model.At the same time,the feature fusion module and the atrous convolution module are adjusted,and the further improved algorithm has stronger detection ability.In this thesis,the further improved algorithms are called Efficent Net B0-Ours+ and Efficent Net B3-Ours+.For Pedestrian detection task for railway scene,both Efficent Net B3-Ours+ and Efficent Net B0-Ours+ which has increased input image resolution can obtain higher detection accuracy when the parameters and calculation amount are far less than SSD.