Neural Architecture Search for In-Vehicle Depth Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In-vehicle computation systems have limited processing power, making it difficult to rapidly identify objects for autonomous driving, especially at higher speeds, as large neural networks with many neurons are unlikely to be processed effectively.
Innovation Solution
The combination of neural architecture search (NAS) and knowledge distillation (KD) is used to generate a smaller neural network model that can be executed by in-vehicle systems for object identification, excluding unnecessary knowledge and focusing on specific task functionality, such as depth analysis of RGB images, to enable rapid object detection and distance estimation without specialized sensors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large neural networks with many neurons are used for object identification, then accuracy of object detection is improved, but processing speed deteriorates due to limited in-vehicle computation power
Solution Approach 1:
The patent extracts and retains only the essential knowledge and functional components needed for object identification, removing unnecessary complexity from large neural networks. This is achieved through knowledge distillation where a compact student network learns only the critical patterns from a larger teacher network, enabling accurate object detection with reduced computational requirements.
Solution Approach 2:
The patent changes the parameters of the neural network by optimizing architecture configurations, reducing the number of neurons and layers while maintaining detection accuracy. Neural architecture search automatically adjusts network parameters to find the optimal balance between model size and performance, enabling faster processing on in-vehicle systems without sacrificing detection accuracy.
2Measurement precision
If large neural networks are used for object identification, then detection accuracy is improved, but device complexity increases making it unsuitable for in-vehicle systems
Solution Approach 1:
The patent extracts only the essential functional components and knowledge patterns from complex large-scale neural networks. Through knowledge distillation, the student network captures critical detection capabilities while eliminating redundant complexity, resulting in a simplified model suitable for in-vehicle deployment that maintains high detection accuracy.
Solution Approach 2:
The patent creates a simplified copy (student network) that replicates the essential functionality of the complex teacher network. This copy contains only the necessary knowledge and patterns for object identification, reducing device complexity while preserving detection performance for in-vehicle applications.
3Reliability
If traditional neural network training is used, then model performance is improved, but development time increases
Solution Approach 1:
The patent performs preliminary actions by pre-training large teacher networks offline and using them to distill knowledge into compact student networks. This preliminary training phase separates the computationally intensive work from the deployment phase, allowing rapid development of optimized models for in-vehicle systems without sacrificing performance.
Solution Approach 2:
The patent introduces knowledge distillation as an intermediary process between traditional training and deployment. The teacher network serves as a mediator that transfers essential knowledge to the student network, enabling efficient model development that maintains high performance while reducing development time through automated architecture search and targeted knowledge transfer.
Data Source
AI summary
An in-vehicle model training system includes a non-transitory computer readable medium for storing instructions; and a processor. The processor is configured to receive an input image; perform object detection, using an encoder, on the input image to identify at least one object, wherein the encoder includes an in-vehicle neural network (NN) model; and generate a first heatmap based on the determined distance to each identified object. The processor is configured to compare the first heatmap with a second heatmap generated by a trained neural network (NN); update the in-vehicle NN model based on differences between the first heatmap and the second heatmap; and determine whether a latency of the encoder satisfies a latency specification. The processor is configured to output the in-vehicle NN model in response to the latency satisfying the latency specification and the difference between the first heatmap and the second heatmap satisfying an accuracy specification.