Distance Object Detection Training Using Aligned Vehicle Sensor Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles equipped with sensors face challenges in accurately detecting objects at a distance due to less precise data acquisition, leading to potential collisions, as the machine learning algorithms (MLAs) are not trained on less accurate representations of objects, resulting in reduced recognition accuracy.
Innovation Solution
A method where a second vehicle follows a first vehicle equipped with similar sensors, allowing both to acquire data simultaneously, which is then aligned and used to train the MLA to recognize objects in less accurate representations, enhancing detection capabilities in different physical settings and sensor types.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Length of stationary object
If sensor data is acquired from a second vehicle at a greater distance, then the detection range is extended, but the measurement precision deteriorates
Solution Approach 1:
The patent uses data from a first vehicle (closer to objects) as a reference copy to train the machine learning algorithm. The MLA learns to map relationships between close-range reference data and far-range sensor data, enabling accurate object detection at distance by copying knowledge from high-precision close-range measurements to low-precision far-range detections
Solution Approach 2:
The patent changes the training parameters of the MLA by introducing distance-dependent training data. The system trains the algorithm using paired data at different distances, allowing the MLA to adapt its detection parameters and compensate for the degradation in measurement precision that occurs at greater distances
2Manufacturing precision
If machine learning algorithm is trained only on close-range data, then the manufacturing precision of detection model is high, but the adaptability to different physical settings deteriorates
Solution Approach 1:
The patent makes the detection model universal by training it on multi-distance data. The MLA learns distance-invariant features and relationships that allow it to function accurately across different physical settings and sensor distances, transforming a specialized close-range detector into a universal detection system that works at various ranges
Solution Approach 2:
The patent performs preliminary training action by exposing the MLA to diverse distance conditions during the training phase. This preliminary exposure to multiple distance scenarios prepares the algorithm to handle various physical settings in deployment, preventing poor performance when encountering distances different from the narrow training distribution
Data Source
AI summary
A method and server for training a machine-learning algorithm (MLA) to detect objects in sensor data acquired by a second sensor mounted on a second vehicle located at a second distance from the objects, the MLA having been trained to detect the objects in sensor data acquired by a first sensor mounted on a first vehicle located at a first distance from the objects. First sensor data acquired by the first sensor on the first vehicle is aligned with second sensor data acquired by the second sensor on the second vehicle. The MLA determines objects and objects classes in the aligned first sensor data. The object classes in the aligned first sensor data are assigned to corresponding portions in the aligned second sensor data. The MLA is trained on the labelled portions in the aligned second sensor data to recognize and classify objects at the second distance.