Video Event Detection Based On Behavior Recognition Of Train Driver

Informatization is both an opportunity and a challenge for the rapid development of Chinese railway.Although CRH train brings people with convenience,the safety problem which faces is also increasingly serious.Especially a series of railway accidents happened in recent years,which again draws public attention to driving security.However,traditional monitoring methods can hardly meet the requirements of rapid response to abnormal events.Therefore,the introduction of the intelligent video surveillance system for railway driver,which has become an urgent demand and an important guarantee for the safety of railway operation.Based on the algorithms of behavior recognition,this paper analyzes the action of train diver,the aim of which is to respond timely to anomalies of record images with the help of machine vision.The system sends a warning message to the terminal and reminds the operator to respond,which will effectively reduce and avoid the occurrence of railway accidents.The achievement of transforming “post-processing” of traditional motoring into an intelligent “pre-processing” will efficiently avoid any potential losses cased by accidents.The main research contents are as follows:(1)This paper introduces the state-of-the-art video event detection and behavior recognition and analyzes the significance of intelligent video surveillance for railway traffic safety,which further illustrates the practical significance of computer vision for safe operation in the field of railway industry.(2)The framework of the monitoring system for railway driver was designed.First,the collected video data is processed for preprocessing in the low-level stage.And then the standards-compliant data will be used as input data,which is processed by the machine vision algorithms,so as to transform low-level image information into high-level semantic information.Finally,video content understanding can be expressed by the semantic description.(3)The recognition steps include object detection,tracking,behavior recognition and semantic description by analyzing the characteristics of driver behavior.Based on the principle of compositionality,an anomaly judgment model was proposed to generate the abnormal warning,which reminds operator to identify the event type by semantic description.In the object detection of middle-level,an object detection algorithm was proposed,which bases on local features and combines the advantages of frame difference and skin color information.In order to distinguish between the head and hands,HOG+SVM was adopted to realize the detection of head.In the event detection of high-level,a new frame was put forward,which modularizes the problem of abnormal detection into a set of variables and constitutes the low-level description of the scene.The experimental results demonstrate the accuracy and real-time of the proposed framework meet the demands of the system.Finally,this paper summarizes the work of the whole thesis and further makes clear the future research direction.