Expanded Bounding Box Tracking for Autonomous Vehicles
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous and semi-autonomous vehicles face challenges in accurately detecting and tracking objects due to varying sensor perceptions and occlusions, leading to inconsistent bounding box sizes and centers, which hinder navigation and collision avoidance.
Innovation Solution
The system employs state estimation using factor graphs to determine the largest bounding box and expand it consistently, with the center of the expanded bounding box serving as the anchor point for tracking, incorporating the object's heading vector and radial direction to resolve ambiguity in expansion direction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the sensor detects objects at different poses over time, then the object detection capability is improved, but the bounding box size and center become inconsistent
Solution Approach 1:
The system performs preliminary actions by detecting the object at multiple poses over time before final tracking. By collecting bounding box data across multiple time steps and poses, the system builds a historical record that enables later determination of a consistent largest bounding box and stable center point, resolving the inconsistency caused by varying detection angles and occlusions
Solution Approach 2:
The system uses feedback by continuously monitoring the object's pose and bounding box characteristics over time. The tracking module feeds back the determined largest bounding box and center to adjust and maintain consistent tracking parameters, allowing the system to compensate for variations in sensor perspective and occlusion conditions while preserving detection accuracy
2Adaptability or versatility
If the sensor perspective changes as the vehicle moves, then the navigation capability is improved, but the perceived object size varies due to occlusion
Solution Approach 1:
The system performs preliminary detection at multiple poses and time steps to capture the object's full extent before navigation decisions are made. By accumulating bounding box data from different vehicle positions and object orientations, the system determines the largest bounding box that represents the complete object size, compensating for occlusions that occur at any single moment
Solution Approach 2:
The system changes parameters by considering multiple bounding box instances with varying sizes and positions over time. Instead of relying on a single bounding box measurement, the system evaluates a sequence of bounding box parameters and selects the largest one, thereby obtaining an accurate object size measurement that accounts for varying sensor perspectives and occlusion conditions
3Measurement precision
If the bounding box is adjusted to fit the detected object at each pose, then the detection accuracy is improved, but the tracking stability deteriorates
Solution Approach 1:
The system performs preliminary adjustments to multiple bounding boxes detected at different poses before finalizing the tracking parameters. By evaluating bounding boxes from a sequence of detections and determining the largest one, the system establishes a stable reference that maintains both detection accuracy and tracking consistency over time
Solution Approach 2:
The system uses feedback mechanisms to maintain tracking stability by continuously comparing new bounding box detections with the established largest bounding box. The tracking module uses the determined center and size as reference parameters, adjusting for variations while maintaining consistency, thereby preserving both detection accuracy and tracking stability
Data Source
AI summary
Systems and methods of tracking objects around an ego vehicle, using state estimation of the objects, where state estimation may include an estimate of a position, velocity, acceleration and bounding box size of the tracked object are disclosed. An object may be detected at one or more poses over time. For each pose, a first bounding box may be generated. The largest bounding box may be determined. For each pose, the first bounding box may be expanded to an expanded bounding box. The expanded bounding box may match the largest bounding box. For each pose, the center of the expanded bounding box may be determined. The object may be tracked according to the center of the expanded bounding box.