Neural Network Object Labeling via Stationary Sensor Confidence
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current convolutional neural networks (CNNs) used in autonomous vehicles face challenges in accurately labeling objects in diverse environmental conditions and backgrounds, leading to inefficiencies in training and limited ability to generalize beyond the training dataset, resulting in false alarms and reduced performance in real-world scenarios.
Innovation Solution
The method involves determining object labels and confidence levels using both vehicle-mounted and stationary sensor images, with the neural network being retrained using higher-confidence labels from stationary sensors when necessary, allowing for improved object detection and vehicle path determination based on polynomial functions to avoid objects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If convolutional neural networks are trained with extensive training data to improve object detection accuracy, then the ability to generalize improves, but the training time and computational resources increase significantly
Solution Approach 1:
The system performs preliminary actions by using stationary sensors to pre-label objects in the training environment before the vehicle encounters them. These pre-labeled objects are then used to quickly retrain or fine-tune the neural network, avoiding the need for extensive traditional training processes while maintaining high detection accuracy.
2Adaptability or versatility
If convolutional neural networks are trained with diverse training data to improve generalization ability, then performance in real-world scenarios improves, but the complexity of data collection and processing increases
Solution Approach 1:
Stationary sensors act as intermediaries between the training environment and the vehicle's neural network. These sensors provide pre-labeled object data that serves as a bridge, enabling the network to learn from diverse real-world scenarios without requiring the vehicle itself to collect and process complex training data.
3Device complexity
If convolutional neural networks use vehicle sensor data alone for object labeling, then the system remains simple, but accuracy in diverse environmental conditions deteriorates
Solution Approach 1:
The system merges data from two sources: vehicle-mounted sensors and stationary sensors. By combining these data sources, the system achieves high object labeling accuracy in diverse environmental conditions while maintaining relative system simplicity through the use of pre-labeled stationary sensor data.
Data Source
AI summary
A computer, including a processor and a memory, the memory including instructions to be executed by the processor to determine a vehicle sensor object label, and a vehicle sensor confidence level for the vehicle sensor object label, by processing a vehicle sensor image with a neural network, determine a stationary sensor object label, and a stationary sensor confidence level for the stationary sensor object label, by processing a stationary sensor image with the neural network, and if the stationary sensor confidence level is greater than the vehicle sensor confidence level, operate a vehicle by determining a vehicle path based on the stationary sensor object label.