Deep Learning Driving Risk Assessment Using Image and CAN Data

The patent replaces traditional rule-based risk assessment mechanisms with deep learning algorithms (CNN and reinforcement learning). The system uses neural networks to automatically process image data from sensors and CAN bus data, substituting manual or simple algorithmic risk evaluation with intelligent systems that continuously learn from data, thereby significantly improving measurement precision while managing complexity through automated processing.

The patent transforms the risk assessment approach by changing from static rule-based parameters to dynamic learned parameters. The deep learning models adapt their parameters (weights and biases) based on training data, allowing the system to capture complex patterns in driver behavior and environmental conditions that traditional fixed parameters cannot detect, thus improving assessment accuracy.

The patent divides the complex deep learning system into two distinct modules: a CNN-based image processing module that extracts visual features from sensor data, and a reinforcement learning module that processes temporal sequences and makes risk predictions. This segmentation allows each module to specialize in specific tasks, improving overall precision while making the system more manageable and interpretable.

The patent introduces an intermediary processing layer that bridges raw sensor data and final risk assessment. The CNN extracts intermediate image features, which are then combined with CAN bus data and processed by the reinforcement learning agent. This intermediary structure simplifies the overall complexity by breaking down the transformation from raw data to risk prediction into manageable stages with clear intermediate representations.

The patent implements preliminary action by pre-training the deep learning models offline with large datasets before deployment. During actual driving, the pre-trained models process images rapidly without requiring extensive real-time computation. The CNN and reinforcement learning agent are prepared in advance to recognize patterns, enabling fast inference while maintaining high accuracy, thus resolving the conflict between processing comprehensive data and minimizing time loss.

The system employs periodic action by processing images at strategically selected time points rather than continuously. The reinforcement learning agent determines when risk assessment is necessary based on contextual cues, processing full image data only when needed while using simpler monitoring between assessments. This periodic processing maintains measurement precision for critical events while significantly reducing overall processing time and computational load.