Driver Abnormal Behavior Recognition Algorithm And System Design Based On Human Segmentation

Driver monitoring system has become a necessary technology to prevent distracted driving and ensure traffic safety.It is mandatory to install driver monitoring system in some vehicle models in China.However,the system still has the problem of low accuracy of abnormal driver behavior recognition algorithm due to the background interference.Therefore,this paper designs the driver abnormal behavior recognition algorithm based on human segmentation and applies it in the driver abnormal behavior recognition system.The main research contents of this paper are as follows:Research on abnormal driver behavior recognition algorithm based on human segmentation.Aiming at the background interference problem of driver abnormal behavior recognition algorithm,this paper proposes a driver abnormal behavior recognition algorithm based on human segmentation.Firstly,a joint network of human pose estimation and human segmentation is proposed to realize the two tasks using one network in driver monitoring system.The speed of this network is better than the combination of existing human pose estimation network and semantic segmentation network.Then,through the method of pre-training and knowledge distillation,the network can achieve similar performance to the large network in the task of driver pose estimation and driver body segmentation,and can accurately remove the background when applied to the task of driver abnormal behavior recognition.After that,in order to identify the abnormal behavior of drivers in the foreground image,the improved VGG16＿DRIVER network based on VGG16 network is proposed,which is faster and more suitable for the task of abnormal behavior identification of drivers.The method of multi-label classification is used to solve the problem of multiple abnormal behaviors occurring simultaneously.Finally,the feature visualization and experimental data analysis were carried out by using the VGG16＿DRIVER network under the condition of preserving and removing the background.The experimental results show that the VGG16＿DRIVER network’s F1-score of the driver abnormal behavior recognition task is increased by 10.83 and 7.80 percentage points respectively in two different background scenes by removing the background,which is better than other existing methods.The design of driver abnormal behavior recognition system and its implementation on embedded platform.In this paper,the proposed driver abnormal behavior recognition algorithm is applied in the driver abnormal behavior recognition system,and the system is deployed in the embedded platform equipped with HI3516DV300 So C chip to realize the real-time detection and voice alarm of the driver abnormal behavior.Synthetically using CPU,neural network acceleration engine and other processing units to achieve system functions.