LiDAR Remainder Classification for Unexplained Object Returns
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicle perception systems face challenges in efficiently updating object tracking and identification due to delays in processing new measurement data, particularly in distinguishing between objects and obscurants like ground, rain, or debris, which requires expensive labeling techniques and is prone to false positives and inaccuracies.
Innovation Solution
A remainder explainer system using a machine learning model partitions sensor data into objects and obscurants, predicting unexplained data as known objects and ignoring the rest, thereby simplifying classification and reducing redundancy, and handles unexplained sensor data without explicit labeling, ensuring no measurement is left behind.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If expensive labeling techniques are used to classify sensor data into objects and obscurants, then classification accuracy improves, but processing cost and time increase
Solution Approach 1:
The patent segments the classification task into two distinct stages: first, a machine learning model performs rapid preliminary classification of sensor data into objects and obscurants; second, a remainder explainer system processes only the unexplained remainder data. This segmentation allows the system to achieve high classification accuracy while minimizing processing time by avoiding expensive labeling techniques for all sensor data.
Solution Approach 2:
The remainder explainer system acts as an intermediary component that bridges the gap between the machine learning model's preliminary classification and the final accurate classification. It processes only the unexplained remainder data, providing a cost-effective intermediate step that improves overall classification accuracy without requiring expensive labeling for all data.
2Reliability
If all sensor data is processed to ensure complete object identification, then detection reliability improves, but system complexity and processing overhead increase
Solution Approach 1:
The patent extracts and processes only the critical remainder data that was not explained by the machine learning model's preliminary classification. By taking out only the unexplained portion of sensor data for further processing, the system maintains high detection reliability while significantly reducing system complexity and processing overhead compared to processing all sensor data.
Solution Approach 2:
The system performs partial processing by applying the remainder explainer only to the unexplained remainder data rather than all sensor data. This partial action approach ensures that critical undetected objects are identified (maintaining reliability) while avoiding the excessive complexity of processing every sensor return through multiple expensive labeling techniques.
3Loss of information
If traditional object detection methods are used to identify all objects, then object identification completeness improves, but processing speed decreases due to redundancy
Solution Approach 1:
The patent inverts the traditional approach by first using a machine learning model to identify obvious objects, then processing only the remainder (unexplained) data to find missed objects. This inverted approach ensures object identification completeness by capturing both detected and missed objects, while improving processing speed by eliminating redundant processing of already-identified objects.
Solution Approach 2:
The machine learning model performs preliminary classification of sensor data into objects and obscurants before the remainder explainer processes the unexplained data. This preliminary action identifies and removes obviously detected objects from further processing, ensuring complete object identification while significantly improving processing speed by avoiding redundant analysis of already-detected objects.
Data Source
AI summary
A method may include obtaining lidar data comprising a plurality of lidar returns from an environment of an autonomous vehicle. The lidar data may be processed with a machine learning model to generate, for the plurality of lidar returns, a plurality of first outputs that each identify a respective lidar return as belonging to an object or non-object and a plurality of second outputs that identify lidar returns belonging to objects as harmful or non-harmful to the autonomous vehicle. A subset of the lidar returns identified as belonging to objects that (i) do not correspond to any of a plurality of pre-classified objects and (ii) were identified as harmful to the autonomous vehicle may be determined. The autonomous vehicle may be controlled based at least in part on the subset of lidar returns.