Object Detection Training Data Alignment for Vehicle Ranging Sensors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural networks for object detection in automated vehicles face errors due to differences in sensor types and positions between specific vehicles and open-source data sets, leading to inaccurate object detection and classification.
Innovation Solution
A system that uses a self-supervised neural network and a clustering technique to determine a first and second set of objects from ranging sensor data, comparing counts to designate the second set as a training data set, and employs iterative processing to exclude unnecessary readings, ensuring accurate object detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a neural network is trained using open-source data sets, then the training process can be simplified and resources saved, but the accuracy of object detection deteriorates due to sensor type and position differences
Solution Approach 1:
The patent creates a synthetic copy of the open-source data set that mimics the sensor characteristics and positioning of the specific vehicle. This copied data set is then used to train the neural network, allowing the system to benefit from pre-trained models while adapting to specific sensor configurations without requiring manual annotation of new data.
Solution Approach 2:
The patent introduces an intermediary transformation layer that converts data from the standard open-source format into a format compatible with the specific vehicle's sensor system. This intermediary process bridges the gap between generic training data and vehicle-specific requirements, enabling accurate object detection without retraining from scratch.
2Measurement precision
If a neural network is retrained for specific vehicle sensor setups, then object detection accuracy improves, but the time and computational resources required increase
Solution Approach 1:
The patent performs preliminary transformation of the open-source data set into vehicle-specific formats before training begins. By pre-processing and adapting the data in advance, the system avoids time-consuming adjustments during the training process itself, thereby reducing overall training time while maintaining accuracy for the specific vehicle configuration.
3Device complexity
If existing neural networks are used without adaptation, then the system complexity is reduced, but detection errors increase due to sensor mismatch
Solution Approach 1:
The patent modifies key parameters of the neural network model to match the specific sensor characteristics of the vehicle. By adjusting parameters such as input dimensionality, feature extraction rates, and processing thresholds, the system maintains relatively simple architecture while achieving high detection reliability tailored to the specific sensor setup.
Data Source
AI summary
A system for producing a data set to train a neural network for object detection can include a processor and a memory. The memory can store an object detection module and a data set determination module. The object detection module can include instructions that cause the processor to: (1) cause the neural network to determine, from data, produced by a ranging sensor system, a first set of objects and (2) cause another object detection system to determine, from the data, a second set of objects. The data set determination module can include instructions that cause the processor to: (2) compare a first count, of objects in the first set, and a second count, of objects in the second set and (2) designate, in response to the first count being different from the second count, the second set as the data set to train the neural network for object detection.