Enhanced Broad-Learning-Based Dangerous Driving Action Recognition on Skeletal Data for Driver Monitoring Systems.

Area of Science:

• Computer Science
• Artificial Intelligence
• Transportation Engineering

Background:

• Traditional Driver Monitoring Systems (DMSs) struggle with real-time performance and robustness on embedded devices.
• Resource-constrained environments necessitate lightweight and efficient driver monitoring solutions.

Purpose of the Study:

• To develop a novel method for recognizing dangerous driving actions using 3D skeletal data.
• To enhance the robustness, real-time performance, and efficiency of Driver Monitoring Systems (DMSs) for embedded applications.

Main Methods:

• Proposed a Graph Spatio-Temporal Feature Representation (GSFR) to select relevant keypoints from 3D skeletal data, reducing complexity.
• Integrated GSFR with a Broad Learning System (BLS), optimized using sparse feature selection and Principal Component Analysis (PCA).
• Implemented a dual smoothing strategy (sliding window and Exponential Moving Average - EMA) to stabilize predictions and minimize noise.

Main Results:

• The proposed GSFR-BLS model demonstrated superior accuracy, efficiency, and robustness compared to existing methods.
• Dynamic keypoint selection in GSFR effectively reduced computational load.
• Optimized BLS and dual smoothing ensured real-time processing and reliable predictions.

Conclusions:

• The GSFR-BLS model is a highly effective solution for recognizing dangerous driving actions.
• The method is well-suited for practical deployment in embedded DMS applications due to its performance and efficiency.
• This research contributes to improving road safety through advanced driver monitoring technology.