Modular Ensemble Neural Network for Real-Time Driver Distraction Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current systems for real-time safety event detection in vehicles face challenges due to high data storage and processing requirements, limited compute capabilities, and the time and cost associated with retraining single-shot event detectors, leading to inefficiencies in detecting distracted states of drivers.
Innovation Solution
Implementing a modular neural network that can be segmented into independent layers for tuning, and an ensemble neural network with multiple models for face detection, hand action classification, head pose classification, and gaze detection, allowing for efficient and accurate detection of distracted states without requiring extensive retraining of the entire model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a single-shot event detector is used for real-time safety event detection, then the system can detect events in real-time, but the compute capabilities are limited and retraining is time and cost-prohibitive
Solution Approach 1:
The patent segments the neural network into multiple independent layers, allowing selective retraining of specific layers rather than the entire model. This reduces compute requirements and training time while maintaining detection reliability.
Solution Approach 2:
The system dynamically determines which neural network layers require retraining based on the specific event type and available compute resources, optimizing the balance between detection accuracy and computational efficiency.
2Device complexity
If a single-shot event detector is used, then the system structure is simpler, but retraining and fine-tuning are limited due to lack of compute capabilities
Solution Approach 1:
By dividing the neural network into trainable layers and fixed layers, the system enables partial retraining with limited compute resources, improving adaptability without requiring complete model retraining.
Solution Approach 2:
Different layers of the neural network have different properties - some are trainable and others are fixed. This local differentiation allows the system to adapt to specific event types while maintaining overall structural simplicity.
3Measurement precision
If the entire neural network is retrained, then detection accuracy improves, but downtime for updates increases
Solution Approach 1:
Segmenting the neural network enables retraining of only necessary layers, significantly reducing update downtime while maintaining detection accuracy through selective fine-tuning.
Solution Approach 2:
Instead of retraining the entire network, the system performs partial retraining on specific layers, which is sufficient to maintain accuracy for specific event types without the time cost of full retraining.
4Measurement precision
If more compute capabilities are allocated for retraining, then model accuracy improves, but cost increases
Solution Approach 1:
Segmenting the training process allows allocation of compute resources only to necessary layers, reducing overall energy consumption while maintaining model accuracy through targeted fine-tuning.
Data Source
AI summary
A vehicle device may execute one or more neural networks (and/or other artificial intelligence), based on input from one or more of the cameras and/or other sensors, to intelligently detect safety events in real-time. The one or more neural networks may be an ensemble neural network that includes neural networks for detecting a head and hand of a user, neural networks for detecting hand actions of the user, neural networks for detecting the head pose of the user, neural networks for predicting an occurrence of an event, and neural networks for predicting a start time and end time of the event. Further, the neural networks can be segmented into a modular neural network based on metadata. The segmentation of the neural network can define a thin layer of the modular neural network to enable independent tuning of the thin layer of the modular neural network.