Research On Distracted Driving Behavior Recognition Algorithm Based On Human Key Point Detection

As an important part of vehicle safety system,distracted driving behavior recognition is of great value.There are two main difficulties in the identification of distracted driving behavior: huge difference within the class and small difference between the class.The existing distracted driving behavior recognition methods start from inside the classification network and improve the recognition accuracy by enhancing the attention to local features.However,the self-learning and self-attention of the recognition network are prone to scene interference,and there are attention biases or improper handling cases.By analyzing the limitations and difficulties of existing distracted driving recognition methods,thesis proposes a two-stage recognition network,introduces the driver’s key point information obtained in the first stage into the classification network,and then carries out the research on distracted driving behavior recognition algorithm based on human key point detection.The main contributions of thesis are as follows:(1)From the perspective of obtaining the region of interest in advance,thesis proposes a two-stage dual-channel distracted driving recognition method.In the first stage,Alphapose key point detection network based on SF3 D data set pre-training was used to obtain driver key point information.In the second stage,Res Net-50 is used as the backbone.In order to give full play to the role of key points,thesis designs two kinds of key region maps by using the key point information,and proposes three kinds of fusion structures in the two-channel fusion stage of the original map and key region map.Starting from inside the network,thesis introduces spatial and channel attention mechanisms to enhance the learning of interested features.(2)On the basis of the existing two-channel network,thesis innovatively introduces the convolutional network model of GCN graph.It is found in the study that the graph structure data formed by connecting drivers’ key points according to the distribution of human bones has a strong correlation with drivers’ distracted driving behavior categories.In view of the shortcomings of GCN network,thesis proposes a collaborative identification method of GCN network and dual-channel network,and adopts two collaborative strategies according to different collaborative modes.The first strategy is to take the two networks as two branches,process the image data and the graph structure data respectively,and fuse them in the prediction layer.In strategy two,a binary classifier composed of 45 GCN binary classification models is constructed and used to correct the top2 results of the two-channel network prediction layer.(3)In thesis,ablation and comparison experiments were designed for the two proposed methods respectively,and the effectiveness of the algorithm was verified from the recognition accuracy of single category and overall category.Compared with the base network model,the accuracy of the two methods proposed in thesis is improved by 2.6% and 3.6% respectively.Compared with other algorithms,the proposed method achieved better recognition effect in 7 categories of SF3 D data set,while poor recognition effect in the remaining 3 categories,with an overall accuracy of 93.2%.