Abstract

Transportation technologies are currently experiencing rapid advancements in the context of global development. The rapid increases higher in traffic volume, however it has contributed to higher, and severity of traffic are continuously rising. The preservation of human lives and attributed to road traffic accident has become vital worldwide and the public is an urgent requirement. However, owing to frequent and widely diverse dangerous driving behaviors of drivers, there are significant potential road safety risks, which will seriously affect the healthy development in the transportation industry. To mitigate such safety hazards, there is an urgent need for precise detection and warning of dangerous driving behaviors of human drivers under driving condition. This dissertation research focuses on a multi-scale data-driven method for detecting drivers’ dangerous driving behaviors. The multi-dimensional data basis provided the essential providing theoretical basis and technical support platform is proposed for formulating of a comprehensive driver warning system.

The main work of this paper includes the following five parts:

(1) Analyzing the danger of a driving behavior and its relationship with traffic accidents from the perspectives of public safety and system engineering is developed considering a visual recognition detection system architecture for dangerous driving behaviors based on using the reported research methods, and deep learning neural network models and detection algorithms used for dangerous driving behavior recognition, laying the theoretical foundation for subsequent research.

(2) An efficient identification method is subsequently proposed for dangerous driving behaviors based on the improved YOLOv8 (You Look Only Once version 8) neural network platform. To improve the recognition efficiency and accuracy of dangerous driving behavior detection, a Multi-Head Self-Attention (MHSA) attention mechanism module is further adopted to enhance efficiency and accuracy, enabling the model focus on global target within a larger receptive field, thereby enhancing the recognition capability of targets. This global modeling capability helps reduce false positive and false negative rates in target detection, improving the overall performance and robustness of the model. Meanwhile, inserting a driver's driving emotion detection module in the network layer shares ROI (Region Of Interest) features in the target detection algorithm, effectively increasing the detection of driver's driving emotion recognition function without significantly increasing the complexity of network computation demand.

(3) Proposing a method for detecting driver's brake pedal and accelerator pedal operations using the Mask-RCNN instance segmentation deep learning network. It recognizes and evaluates the operation of driver's brake pedal and accelerator pedal through video frames collected using a driver's leg camera. The use of ROI Align method to effectively align pixels during downsampling enhances the feature extraction capability of the entire detection network with increasing the computational efficiency and complexity of the network model, thereby improving the efficiency and accuracy of the detecting driver's brake pedal and accelerator pedal operations.

(4) A data-driven detection method is subsequently proposed using the filtering and sliding windows. By collecting brake signals and throttle signals from the vehicle CAN bus data and performing low-pass and median filtering, the computational feature extraction capability of the driving signal data is enhanced. The sliding window method compares and judges brake signals and throttle signals in the sampling interval with the corresponding thresholds, achieving detection and evaluation of driver's dangerous driving behaviors such as sudden acceleration and deceleration.

(5) Building a dangerous driving behavior detection system, integrating neural network training weights and detection algorithms through Python and Qt Designer software systems, and designing a visual real-time detection front-end UI interface. At the same time, designing real vehicle verification experiments, based on real-time experiments in actual driving environments, to verify the effectiveness and superiority of the driver's dangerous driving behavior detection method proposed in this study, providing theoretical basis and technical support for the widespread application of dangerous driving behavior recognition technology.