Research And Implementation Of Driver Behavior Recognition Based On Deep Neural Network

Driver behavior recognition is an application scenario of human behavior recognition.During driving,whether the driver’s driving behavior meets the safety standards is directly related to the personal safety of the entire vehicle,so video surveillance for drivers is an important security measure.In traditional methods,professionals are required to manually design target features and extract them through specific methods.Because of its self-learning characteristics,deep neural networks have more freedom to automatically extract valid information from the data.The combination of deep learning and human behavior recognition has driven the development of related technologies and applications.Mainly focusing on the application scenario of bus drivers’ driving behavior,this thesis carries out research and implementation of driver behavior identification by analyzing the real video data collected from real shots,performing necessary preprocessing on the data,and applying deep neural network model to extract data features and identifies data categories.The main work includes:(1)Filter the video for the speed of the bus when recording.Through the subtitle information in the video picture,the template matching method is used to identify the driving speed when recording,and the video whose speed is lower than the threshold does not continue to be detected.(2)Propose a scheme for locating the driver by detecting the steering wheel.Through experimental analysis,it is found that the relative position between the steering wheel and the driver in the picture is stable.The steering wheel has the advantages of small target and prominent arc features.Using histogram of oriented gradient as feature descriptor and support vector machine as the classification model,the steering wheel area can be well detected by multi-scale detection algorithm with extra filtering,and then the driver screen area can be obtained by extending the steering wheel area.(3)Propose a deep VGG16-BiLSTM driver behavior recognition model,which uses a deep convolutional neural network to extract the characteristics of the frame picture,and sends the features into a Bi LSTM model in order to identify the category of the video.Experiments show that the model presented in this thesis is deep both temporally and spatially,and has a better recognition effect than shallower recognition models.In order to meet the needs of the special indicators,false-alarm rate,of the project,this thesis thresholds the output vector of the model and obtains a decrease in the false-alarm rate.(4)Based on the above research and work,a set of driver violation detection system was implemented,which is applicable to the method of sampling and inspecting car monitoring videos,centralized on the remote monitoring platform to detect the driver’s behavior in non-real time,which can reduce the production cost of the front-end camera hardware and reduce the promotion difficulty.Through theoretical analysis and experimental verification,the system can achieve a good recognition of the video submitted for inspection.It can be used to supervise drivers when applied to vehicle-mounted video surveillance,and the intelligent identification can help reduce labor on the monitoring platform.