Research On Complex Behavior Analysis Of Train Drivers Based On Active Learning

In recent years,with the rapid development of China’s railroad industry,people are increasingly inclined to choose trains as the main mode of travel.In order to protect the safety of passengers on the way,the behaviors of train drivers needs to be regulated and restrained.At present,many relevant departments of the railroads still use manual viewing to verify the monitoring video of train cabs as a way to regulate and restrain the behaviors of train drivers,and this method often requires a lot of human resources and time costs.Therefore,many researchers have turned their attention to the field of machine learning,aiming to intelligent analyze the driver’s behaviors through advanced computer vision technology as a way to restrain and regulate the driver’s behaviors.However,in the process of solving real-world problems,they often face problems such as insufficient effective samples,unbalanced data,complex real-world scenarios,and poor model generalization ability.To address the above problems and challenges,this paper designs and develops an intelligent video analysis system for train drivers based on active learning theory and related technologies,and the main research contents of the paper are as follows.(1)An adaptive category suppression adversarial active learning method is designed to filter information-rich data by active learning to achieve a deep learning model with superior performance using a small amount of data training.The deep learning model with superior performance is obtained.Meanwhile,two independent adaptive category suppression object classifiers are constructed based on the adaptive class suppression loss detection algorithm ACSL,which uses ResNet-50 as the backbone network and uses the adaptive category suppression loss function to replace the original common cross-entropy loss function,further solving the problem of longtail data in the active learning object detection task.(2)The train driver dataset was designed and constructed by combining the behavioral guidelines of train drivers in actual operations,while data augmentation was performed for samples with a small number of categories.At present,there are many drawbacks in the way of large-scale manual labeling of data,and this paper proposes an automatic labeling method for the dataset based on active learning technology.Highvalue data are first obtained by active learning filtering strategy,and then they are automatically annotated.In the process of automatic dataset annotation,only manual review of the automatically annotated data is required,which saves a lot of annotation costs.(3)Based on the above work,an intelligent video analysis system for train drivers is designed and developed.The system combines speed constraint,time constraint and space constraint and uses Faster R-CNN detection network to analyze the train driver video.The system has been successfully applied to cab video analysis tasks in several railroad bureaus and has achieved good social and economic benefits.