| The use of intelligent vehicle assistance systems to identify the driver’s driving intentions in advance to reduce accidents has been a hot topic in the field of automotive safety.Currently,driving intentions can be decoded from surface electromyography(s EMG)of the lower limbs using artificial intelligence algorithms.But these methods do not consider the correlation between muscles(i.e.the relationship between muscles during braking driving actions).In this paper,the s EMG of drivers during different braking processes are collected to construct connection graph of inter-muscle signals.And they are used as inputs to graph neural network algorithms to identify drivers’ braking intentions.The research is as follows.(1)Construction and analysis of the connection graphs of lower limb musclesThe experimental scheme of driving intention was designed to collect multi-channel s EMG from drivers during regular braking and emergency braking in various types of driving.Using correlation coefficients and maximum information coefficients to explore the correlation,which between the channels of s EMG of lower limb actions.And the method can be used to construct the connection graphs.The topology of the muscle connection graph was analyzed using node characteristics and importance indexes to evaluate the importance of each node.The results of the analysis were used for the identification of conventional driving and braked driving.(2)Construction of a lower limb action intention recognition model based on graph neural networkBased on the muscle connection graph lower limb muscles,an end-to-end lower limb action intention recognition model is constructed using graph neural network.The algorithm can simultaneously learn the temporal and spatial characteristics of each s EMG channel.The model recognized six different lower limb actions in the open-source lower limb motion dataset with an accuracy of 0.954.On this basis,the improved model using an adaptive graph construction method,and the accuracy is improved to 0.962.The accuracy of continuous actions(including transitional states of action changes)in 17 types is significantly higher than that of other models,with an accuracy of up to 0.826,which provided a new method for subsequent braking driving recognition of lower limbs.(3)Construction of a lower limb driving recognition model based on graph neural networksUsing the braking driving dataset,the model structure of the lower limb action intention recognition method,which based on graph neural network,was optimized to construct a lower limb driving braking intention recognition model.The accuracy of the model for braking and emergency braking action recognition reached 0.895,which is better than other algorithms under the same conditions.The research provided a new method and idea for the driving intentions recognition.In summary,this paper addresses the problem of driver braking intention recognition,analyses the muscle association of the action occurrence process through the driver’s surface EMG signals,constructs a muscle connection graph for the corresponding action,and designs a motion intention recognition model based on graph neural networks,obtaining a better recognition accuracy in the public dataset of lower limb motion and the home-made braking dataset. |
