Automatic Recognition And Analysis Of Train Driver Working Conditions

Safety has always been a top priority in China's railway operations,and in order to achieve this goal,China railway has developed a set of working specifications for drivers in railway operation.During the running of the train,the train driver must stay focused at all times.When he encounters the signal of railway running or receives the command from the control center,he must immediately make corresponding operation or gesture to ensure the safety and normal running of the train.At present,most of the supervision of the working status of the driver during the running of the train consists of the staff checking the working status of the driver through the monitoring video of the cockpit of the train and the rapid scanning of human eyes,which has caused a huge waste of resources in terms of manpower and material resources.Based on the research in the process of the train driver characteristics of all kinds of working state,decided to adopt deep learning algorithm for automatic identification,the status of the driver from the selection of recognition algorithm,data set of production,recognition model training and optimizing the classification accuracy of a complete and detailed analysis,this paper aims to address the current supervise drivers working state by artificial existence problem of waste of resources and low rate of fault tolerance.The main research contents and work arrangement of this paper are as follows:Choose the appropriate deep learning algorithm.First of all,the three algorithms of YOLO,Deep Pose and dual-stream network are introduced in detail,including the implementation principle,processing speed and recognition effect of the algorithm.Then analyze the advantages and disadvantages of the three deep learning algorithms from different angles,combined with the characteristics and application scenarios of the train driver's working state in this article,and finally choose the YOLO algorithm that is simple to implement,fast,and has a good recognition effect.Dataset production.This paper designs and recognizes the six working states of the driver,namely "leave" state,"work" state,"forword" state,"swing" state,"wave" state and "call" state.Through the analysis of the specific characteristics of the driver's working state in the surveillance video,a method of making an "end-to-end" data set based on the KCF algorithm is proposed,which directly tracks the categories and marks on the surveillance video and outputs the data for model training.Samples and corresponding annotation information files.For the generated data set,it is necessary to check the correctness of the labeling frames and labeling information of each category in the sample to ensure the availability of the data set.Model training.Select the Tensor Flow deep learning framework to build the network structure of the model,design the loss function of the network in the form of mean square error and binary cross entropy,iteratively update the network parameters through the Adam optimization strategy,and use GPU acceleration to complete the training of the network model under window.During the model training process,the model training situation is visualized through average loss data.After the model training is completed,the model is tested,and the advantages and disadvantages of the model are analyzed from the four indicators of precision,recall,average IOU,and m AP.Model optimization.To solve the problem of unsatisfactory recognition effect caused by the imbalance of the number of categories of the data set in the model,through the analysis of the data set,adjust the proportion of each category in the number of training data sets,and modify the loss function of the original model to Focal The loss form strengthens the model's ability to learn difficult-to-resolve categories,thereby improving the model's recognition effect.Through the analysis and research of the above content,the resulting model performs well on the test set,has the characteristics of fast recognition speed and high recognition accuracy,and can achieve automatic recognition of the working state of the train driver.