Research On Driving Behavior Recognition Algorithms Based On Computer Vision

In recent years,the increasing traffic technology has changed the way people travel.However,frequent traffic accidents not only seriously harm the lives of the public,but also often cause great economic losses.According to the analysis,most traffic accidents are caused by the driver’s fatigue driving or distracted driving,which leads to the decrease of vehicle handling ability.In order to reduce the accident rate caused by human,when the driver is in the state of fatigue or distraction,the driver’s state and behavior are detected and given corresponding warnings,which can effectively improve the driver’s safety awareness and standardize the driving behavior.Based on this,this paper aims to propose an efficient and accurate driver behavior recognition algorithm to achieve real-time monitoring of driving behavior.On the basis of investigating the related researches at home and abroad,this thesis uses two technologies to realize the effective recognition of driving behavior.The first one is the behavior recognition scheme from the whole to the part based on face analysis,the second one is the end to end driving behavior recognition scheme based on deep learning.Aiming at the driving behavior recognition algorithm based on facial analysis,firstly,the driver’s face location is realized efficiently and accurately by using the face tracking module and the face quadratic discriminant module;secondly,the driver’s fatigue detection,phone call detection and smoking detection in real scenes are realized by dividing the detection sensitive areas and constructing corresponding event detectors.The algorithm achieves real-time and accurate effects.Aiming at the driving behavior detection algorithm based on deep learning,considering the difference between driving action and locality of driving action,three recognition models with higher accuracy are proposed.(1)The multi-branch attention convolution neural network model is based on multiscale features,and the spatial attention mechanism is introduced to enhance the different scale features representation.Experiments show that the proposed algorithm has obvious improvement effect.(2)Densely channel to spatial attention transmission network is designed considering the local characteristics of driving action.The channel features are corrected through the channel attention model,and then the channel feature maps are encoded to the a spatial mask.Finally,the shallow feature maps are modified with a spatial mask to enhance the details of local action.Experiments show that this method has the ability to focus the key action areas step by step.(3)On the basis of the above model,the two tasks of action location and action classification are decoupled from a network,and the two tasks are coordinated by constructing a tutor-student network of guidance-learning mode.The tutor network provides the location of the action area,and the student network learns the action features according to the action location to obtain knowledge.Experiments show that the model can significantly improve the accuracy of behavior recognition.