Cross-Vehicle Object Label Transfer for Long-Range AV Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for generating training data for autonomous vehicle object detection models face challenges such as the unreliability of simulated RADAR data, limitations of LIDAR at long ranges, inefficiencies in human labeling, and the lack of annotated occluded objects, leading to time-consuming and inaccurate training processes.
Innovation Solution
The method involves generating a long-range objects dataset using proximal time-synced data from multiple autonomous vehicles with intersecting fields of view, leveraging RADAR data to supplement LIDAR limitations by transferring and updating object detections across vehicles, thereby enhancing the accuracy and efficiency of object labeling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If simulated RADAR data is used for training, then training efficiency is improved, but data reliability deteriorates
Solution Approach 1:
The patent uses LIDAR data as an intermediary to transfer object detection information between vehicles. Instead of relying on unreliable simulated RADAR data, the system transfers validated LIDAR-based object detections from one vehicle to another, using the reliable LIDAR data as a mediator to supplement training datasets with accurate long-range object information.
2Measurement precision
If LIDAR is used for long-range object detection, then measurement precision is improved, but detection capability deteriorates at long ranges
Solution Approach 1:
The patent combines LIDAR data from multiple vehicles with intersecting fields of view to overcome the limited long-range detection capability of individual LIDAR systems. By merging detections from multiple sources and transferring object information between vehicles, the system achieves extended long-range detection coverage while maintaining LIDAR's measurement precision.
3Measurement precision
If human labeling is used for object detection training, then labeling accuracy is improved, but processing time deteriorates
Solution Approach 1:
The patent implements a self-service mechanism where object detections from one vehicle automatically supplement the training data of another vehicle through data transfer. This automated cross-vehicle data sharing eliminates the need for time-consuming manual labeling while maintaining high accuracy, as the transferred detections are based on reliable LIDAR measurements.
4Device complexity
If single-vehicle data is used for training, then data processing complexity is reduced, but dataset completeness deteriorates
Solution Approach 1:
The patent adds a spatial dimension to data collection by utilizing multiple vehicles with different positions and fields of view. Instead of relying on a single vehicle's limited perspective, the system aggregates data from multiple spatial sources, transferring object detections across vehicles to create a more complete training dataset that covers a broader range of scenarios and object types.
Data Source
AI summary
Disclosed are embodiments for generating long-range objects dataset using proximal time-synced data for object detection models. In some aspects, a method includes receiving a set of sensor data for a first autonomous vehicle (AV) corresponding to first scene data that is part of an intersecting field of view (FoV) with second scene data of a second AV; inputting the set of sensor data to a simulated trained object detection model of the first AV to generate a first set of object detections for the first scene data; comparing the first set with a second set of object detections generated for the second scene data; supplementing the first set of object detections with information from the second set of object detections to generate a supplemental set of object detections; and generating, using the supplemental set of object detections, a training data set for an object detection model of the first AV.