A Real-Time Driver Behavior Monitoring and Feedback Framework to deliver safe, comfortable, and energy-efficient driving

Understanding and responding to driver behavior is significant for improving road traffic safety, optimizing energy efficiency, and enhancing ride comfort, especially in electric vehicles. This thesis proposes a human-machine interface (HMI) system framework that recognizes multiple types of abnormal driving behaviors in real time and provides personalized and immediate feedback to drivers through an intuitive interface. First, typical roads and traffic scenes are selected using the SCANeR Studio virtual driving simulation platform, allowing realistic driving simulation and data acquisition. Dynamic parameters such as vehicle speed, longitudinal and lateral acceleration, steering angle, pedal position, and motor efficiency are collected. The K-means algorithm is applied for preliminary data segmentation, then the improved DBSCAN (iterative DBSCAN) algorithm is used for unsupervised clustering to identify potential aggressive driving behaviors. These clustering results are further subdivided based on the thresholds of comfort and energy efficiency indicators (e.g., jerk, motor efficiency), resulting in a multi-category label set. A long short-term memory (LSTM) neural network is used to train time series driving data using this label. Bayesian optimization is used to tune hyperparameters such as learning rate and the number of hidden layers. The final model classifies four behavior types in real time: aggressive, uncomfortable, low energy efficiency, and normal, achieving F1 scores of 0.93, 0.96, and 0.98, respectively. Based on this classification model, an HMI is designed and implemented. In addition to presenting real-time dynamic vehicle data, it communicates driving behavior classification results to the driver through visual and auditory alerts. A unified “sensitivity threshold” is introduced, allowing drivers to adjust this threshold through the interface. During operation, the LSTM model continuously outputs probabilities for each behavior type. The system compares each probability with the user-defined threshold in order of priority (aggressive > low energy efficiency > uncomfortable > normal ). Once any behavior’s probability exceeds the threshold, it is regarded as the current driving state, and the corresponding prompt is triggered. This single-threshold, priority-based scheme enhances interaction efficiency and reduces distraction caused by overlapping alerts. The framework is deployed on a real-time, hardware-in-the-loop platform composed of a driving simulator, the Speedgoat, and the Raspberry Pi. The User Datagram Protocol (UDP) is used to enable low-latency communication between modules, ensuring real-time, closed-loop control of the data flow from acquisition, identification, and feedback.