Lightweight Driver Behavior Identification Model with Sparse Learning on In-Vehicle CAN-BUS Sensor Data.

Area of Science:

• Computer Science
• Artificial Intelligence
• Automotive Engineering

Background:

• Driver behavior identification is crucial for connected car safety and functionality.
• Existing methods often require significant computational resources and large datasets.
• There is a need for efficient, lightweight models for embedded automotive systems.

Purpose of the Study:

• To develop a lightweight, end-to-end deep learning framework for driver behavior identification.
• To utilize in-vehicle Controller Area Network (CAN-BUS) sensor data for behavior profiling.
• To achieve superior performance with reduced computational and memory footprints.

Main Methods:

• Proposed a deep learning architecture featuring depth-wise convolution and augmented recurrent neural networks (LSTM/GRU).
• Employed time-series classification on CAN-BUS data with a reduced minimum time-step length.
• Utilized channel pruning for model compression and sparse-learning techniques for adaptability to new classes.

Main Results:

• The proposed method significantly outperforms state-of-the-art driver behavior profiling models.
• Achieved reduced floating-point operations, fewer parameters, compact model size, and less inference time.
• Demonstrated successful deployment on embedded systems (Xavier, TX2, Nano) using NVIDIA Docker.

Conclusions:

• The developed framework provides an efficient and effective solution for driver behavior identification in connected cars.
• The model's lightweight nature and adaptability make it suitable for resource-constrained embedded automotive environments.
• This research advances the integration of AI for enhanced vehicle safety and personalized driving experiences.