Research On Detection Method Of Distracted Driving Behavior Based On Deep Learning

With the rapid development of the economy,the amount of car maintenance in China has increased rapidly,and car travel has become an indispensable part of lives of people.However,this has also brought serious traffic problems.The occurrence of traffic accidents has caused great harm to both the victims and the parties involved,as well as huge economic losses to the country.According to research,traffic accidents are mainly caused by dangerous driving,and distracted driving in dangerous driving is a common behavior in dangerous driving and an important cause of traffic accidents.Therefore,in order to reduce the occurrence of traffic accidents,reduce distracted driving by drivers in the vehicle,and improve driver focus is a meaningful and necessary thing.Nowadays,mainstream distracted driving detection methods have the following shortcomings: First,the model is large,the detection speed is slow,and the deployment is difficult,which cannot meet the requirements of real-time and cost;The second is that in the actual driving environment,there are many low brightness environments,and the detection and recognition rate of distraction behavior in low brightness scenes is low.There are many cases of false detection and missed detection,which cannot meet the detection accuracy requirements.Starting from the above two aspects,this paper proposes a deep learning based distracted driving detection and low brightness image enhancement method in the vehicle,which achieves fast and accurate detection of distracted driving behavior in the vehicle,and can maintain a high recognition rate under low brightness.The method in this paper can accurately detect distracted driving behavior of drivers,which is of great significance in reducing the occurrence of traffic accidents.The main contributions of this article are as follows:(1)This paper analyzes the characteristics of distracted driving detection behavior,and proposes a distracted driving behavior detection algorithm based on improved YOLOv5 s,which realizes the detection of distracted driving behavior.In order to improve the inference speed and recognition accuracy of the YOLOv5 s model,a CBAM attention mechanism is added at the YOLOv5 s model,which strengthens the feature extraction ability of the YOLOv5 s model for key image information and reduces the model weight of non-important image information.In order to adapt the training anchor frame of YOLOv5 s to the detection target in the actual dataset,and solve the problem that the size of the anchor frame does not match and affect the training speed and accuracy,the Kmeans++ algorithm is used to cluster the dataset anchor frame,that is,the size of the anchor frame is adjusted to make it more suitable for the distracted driving dataset in the car in this paper.In order to select the appropriate model loss function,the performance of the YOLOv5 s model under four different loss functions is tested,and the DIOU loss function suitable for the dataset in this paper is selected.After comparison with similar model algorithms,the average detection accuracy(m AP)of the proposed algorithm reaches 98.7%,the model size is only 14.4MB,and the model detection speed reaches 149 FPS,which can meet the task of detecting distracted driving behavior.(2)The influence of low-brightness scenes on distracted driving detection behavior in actual driving is analyzed,and after combining with actual driving scenarios and referring to the characteristics of various mainstream low-brightness enhancement algorithms,an in-vehicle low-brightness enhancement algorithm based on zero reference depth curve is used to enhance the in-vehicle low-brightness environment.The algorithm contains a new type of light enhancement curve(LE-curve)and a depth curve prediction network(DCE-Net)that trains the curve parameters,and the low-brightness image can directly generate the light-enhanced image through the light enhancement curve.After collecting images of low-brightness in-car scenes,it is experimentally proved that under the same experimental conditions,compared with the mainstream deep learning brightness enhancement algorithm,the proposed algorithm has a significant brightness improvement,the fastest recognition speed,the detection speed reaches 1000 FPS,and the time to improve the brightness of an image only takes 0.001 seconds.Finally,on the low-brightness in-car dataset collected by ourselves,the average accuracy(m AP)of the distracted driving detection model was 70.2%,and the average accuracy(m AP)of the distracted driving detection model after low-brightness enhancement was 84.2%,compared with the distraction detection accuracy increased by 14% after brightness improvement,and the improvement effect was significant.(3)The deployment feasibility of the distracted driving detection model is verified,and the distracted driving detection model is deployed to the mobile phone to detect the distracted behavior by using the mobile phone camera.Firstly,this paper converts the distracted driving detection model and model weight conversion on the PC side,and converts the PC model file into the distracted driving detection model file under the NCNN framework.Secondly,the Android Studio mobile application development software is used to convert the distracted driving detection model into a mobile APP and install it in the mobile phone,and finally after the mobile phone detection test,the mobile phone detection model is successfully deployed,and it can run successfully on the mobile phone platform and complete the detection.