Pedestrian-Object Segmentation for Rare AV Perception Cases
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous vehicle perception systems struggle to accurately detect and classify rare objects, such as pedestrians with associated objects, which can lead to reduced safety and navigation accuracy due to incomplete or inaccurate information.
Innovation Solution
A machine-trained model using a top-down segmentation algorithm is trained to detect pedestrian objects by associating bounding regions with pedestrians and their associated objects, enabling the vehicle to determine their motion and geometry, thereby improving object detection and classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If standard object detection algorithms are used, then detection speed is maintained, but detection accuracy for rare objects deteriorates
Solution Approach 1:
The patent applies segmentation by dividing the detection task into multiple stages: first detecting pedestrians, then identifying associated objects, and finally classifying object types. This multi-level segmentation approach improves detection accuracy for rare objects while managing system complexity through modular processing steps.
Solution Approach 2:
The system performs preliminary detection of pedestrians before detecting associated objects. By first identifying pedestrians and then searching for objects in their vicinity, the system prepares detection data in advance, improving overall detection accuracy for rare pedestrian-object combinations.
2Reliability
If comprehensive object detection is performed, then safety information completeness is improved, but processing time increases
Solution Approach 1:
The patent applies local quality by focusing detection resources on specific regions of interest - primarily areas around detected pedestrians. Instead of uniformly processing the entire scene, the system concentrates computational effort where pedestrian objects are likely to occur, improving safety information completeness while reducing overall processing time.
Solution Approach 2:
The system performs partial detection by initially identifying pedestrians and then selectively detecting objects only in their proximity. This partial action approach provides sufficient safety information for navigation decisions without requiring exhaustive detection of all objects in the environment, thus reducing processing time while maintaining reliability.
3Measurement precision
If pedestrian objects are detected with bounding regions, then object position precision is improved, but classification accuracy for rare objects deteriorates
Solution Approach 1:
The patent transitions from two-dimensional bounding box detection to three-dimensional object detection and classification. By adding the dimension of object type classification alongside position detection, the system recovers classification information that would otherwise be lost, enabling both precise positioning and accurate identification of rare pedestrian objects.
Solution Approach 2:
The system uses feedback loops where detection results are continuously refined through multiple processing stages. Initial bounding region detection provides position information, which then feeds into classification algorithms that use this spatial context to improve object type identification, thereby recovering classification information while maintaining position precision.
Data Source
AI summary
Detecting pedestrians with objects is described (pushing stroller, carts, opening doors, carrying umbrellas, etc.). In an example, a perception component of a vehicle can receive sensor data from sensor(s) associated with the vehicle. The perception component can determine, by a model, observation(s) associated with the sensor data, wherein an observation comprises a first object (e.g., a pedestrian). The perception component can determine whether the first object is associated with a second object (e.g., a pedestrian object), wherein the first object and the second object are associated with a compound object (e.g., a pedestrian/pedestrian object system). The perception component can provide the indication of the first object or an indication of the compound object to at least one of a prediction component or a planning component of the vehicle for controlling the vehicle.