Amodal Cuboid Placement for Stable Static Object Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current amodal cuboid algorithms for object detection in autonomous vehicles face issues with stationary objects, particularly when they are partially occluded, leading to 'cuboid jumping' and instability in object tracking, which affects downstream consumers and can result in unsafe vehicle behavior.
Innovation Solution
A novel amodal cuboid-based algorithm that involves obtaining loose-fit cuboids overlaid on 3D graphs to encompass lidar data points, checking for static objects through cuboid dimension, centroid, and rotation checks, and identifying the center of the amodal cuboid to stabilize object tracking and control operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If amodal cuboid algorithms are used for object detection, then object tracking is enabled, but cuboid jumping occurs for stationary objects leading to tracking instability
Solution Approach 1:
The system performs preliminary classification of objects as static or dynamic using historical trajectory data before applying cuboid placement algorithms. This preliminary action allows the system to apply different cuboid placement strategies appropriate to each object type, preventing cuboid jumping for static objects while maintaining tracking flexibility for dynamic objects.
Solution Approach 2:
The system applies different cuboid placement rules based on the local characteristics of each object. For static objects, the cuboid center is constrained to remain fixed relative to the object's detected position. For dynamic objects, the cuboid can adjust its position more freely. This local differentiation resolves the contradiction by providing stability where needed while preserving tracking adaptability where required.
2Loss of information
If amodal cuboids are placed to encompass occluded objects, then complete object representation is achieved, but precision in locating the true object center deteriorates
Solution Approach 1:
The system segments the object detection problem into two independent tasks: (1) determining the amodal cuboid boundaries that encompass the complete object including occluded portions, and (2) separately determining the center position based on visible features and trajectory inference. This segmentation allows each task to be optimized independently, resolving the contradiction between complete representation and precise center localization.
Solution Approach 2:
The system introduces an intermediary computational step that uses historical trajectory data and static object classification as a mediator between the detected cuboid boundaries and the final center position. This intermediary process infers the true center position even when direct observation is blocked, maintaining measurement precision while preserving object completeness.
3Quantity of substance
If loose-fit cuboids are used to encompass lidar data points, then all data points are captured, but precision in object boundary definition is reduced
Solution Approach 1:
The system employs a dynamic, two-stage cuboid fitting process. First, loose-fit cuboids are used to capture all lidar data points ensuring complete coverage. Then, in a second refinement stage, the cuboid boundaries are tightened and adjusted based on the classified object type and trajectory information. This dynamic two-stage approach resolves the contradiction by initially prioritizing coverage and then optimizing for precision.
Data Source
AI summary
Disclosed herein are systems, methods, and computer program products for controlling a mobile platform. The methods comprise: obtaining loose-fit cuboids overlaid on 3D graphs so as to each encompass lidar data points associated with an object; defining an amodal cuboid based on the loose-fit cuboids; checking whether or not the object is static through a time period; identifying a center for the amodal cuboid based on the checking; and causing operations of the mobile platform to be controlled based on the amodal cuboid having the center which was identified.