LRCN Steering Control for Autonomous Driving Power Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional advanced driver assistance systems (ADAS) for autonomous driving require substantial processing power due to the use of very deep neural networks (DNNs), leading to increased costs and power consumption, which affects vehicle efficiency.
Innovation Solution
A long-term recurrent convolutional network (LRCN) model is trained using data from controlled test driving by human drivers to plan and control a sequence of steering wheel angles, eliminating the need for high-powered CPUs or GPUs by utilizing a set of sensors to determine vehicle parameters and mathematical models for verification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If very deep neural networks are used for autonomous driving control, then the driving accuracy and reliability are improved, but the processing power requirements and power consumption increase substantially
Solution Approach 1:
The patent transforms the neural network architecture from a spatially deep structure to a temporally deep structure using recurrence. The LRCN model uses a convolutional network for spatial feature extraction combined with recurrent layers that process temporal sequences of inputs, allowing the system to maintain high reliability through extended temporal context without requiring proportionally deeper spatial networks that would consume more power.
2Measurement precision
If very deep neural networks are used for autonomous driving control, then the driving accuracy is improved, but the device complexity and cost increase
Solution Approach 1:
The patent segments the processing architecture into distinct functional components: a convolutional neural network for spatial feature extraction from sensor inputs, recurrent neural network layers for temporal sequence processing, and a verification module using mathematical models. This segmentation allows each component to be optimized independently and executed on standard processors rather than requiring a single monolithic high-performance computing system.
Solution Approach 2:
The patent introduces mathematical models as an intermediary verification layer between the neural network output and the final control execution. These models provide a lightweight sanity check that ensures the neural network outputs are physically plausible without requiring the neural network itself to be excessively complex, thereby improving precision while controlling overall system complexity.
Data Source
AI summary
Systems and methods for planning and executing a sequence of lateral controls for a vehicle comprise determining a set of parameters comprising road lane lines, a pose of the vehicle, a movement state of the vehicle, and actions by a driver of the vehicle and, while operating the vehicle in an autonomous driving mode, obtaining a long-term recurrent convolutional network (LRCN) model that has been trained with training data captured during controlled test driving by a human driver, using the LRCN model and the set of parameters as inputs, determining a sequence of a plurality of desired steering wheel angles for the vehicle, and when the sequence of the plurality of desired steering wheel angles is verified via comparison to one or more mathematical models, controlling steering of the vehicle according to the sequence of a plurality of desired steering wheel angles.