Object Detection Fusion Using Anchor-Dependent Quality Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection methods using Region Proposal Networks (RPNs) suffer from information loss due to redundant hypothesis discarding and lack of modeling knowledge in post-processing, leading to suboptimal object detection quality.
Innovation Solution
A method that processes sensor signals using RPNs to select and fuse object hypotheses based on an anchor-dependent quality model and target function, identifying redundant hypotheses by proximity to the selected hypothesis, and applying fusion mechanisms like weighted least squares to improve detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If non-maximum suppression is used to select object hypotheses with highest objectness score, then false positives are reduced, but information is lost due to discarding redundant hypotheses
Solution Approach 1:
The patent recovers information that would otherwise be discarded by NMS. Instead of simply discarding redundant hypotheses, the method integrates their bounding box information through fusion operations, thereby recovering valuable spatial data while still eliminating false positives through the quality model-based selection process
Solution Approach 2:
The patent merges redundant object hypotheses by fusing their bounding boxes using integration operations. Rather than selecting only one hypothesis per object, the method combines information from multiple hypotheses to create a more accurate and comprehensive object detection result
2Productivity
If classical post-processing is used to identify redundant hypotheses, then processing speed is maintained, but modeling knowledge of RPN output structure is not utilized
Solution Approach 1:
The patent incorporates feedback by using the known target function structure of the RPN to guide the identification of redundant hypotheses. The quality model leverages feedback about the expected output structure (bounding box relationships to anchor positions) to more accurately identify which hypotheses are redundant, improving identification precision while maintaining efficient processing
3Ease of manufacture
If hyperparameter-based quality assessment is used, then implementation is simple, but detection quality is suboptimal due to lack of anchor-dependent evaluation
Solution Approach 1:
The patent applies local quality assessment by evaluating each object hypothesis based on its specific anchor position and bounding box characteristics. Instead of using a single global hyperparameter, the quality model computes anchor-dependent quality measures that reflect the local geometric relationships, thereby improving detection quality while maintaining computational efficiency through the structured approach
Data Source
AI summary
A method for the detection of an object in an environment of a vehicle as a function of sensor signals of a sensor for acquiring the environment of the vehicle. The method includes: processing the sensor signals using a region proposal network to obtain at least one object hypothesis per anchor, the object hypothesis including an object probability and a bounding box; selecting the best object hypothesis on the basis of a quality model, the quality model being a function of the anchor and the bounding box of the object hypothesis; identifying redundant object hypotheses relative to the selected object hypothesis, the redundant object hypotheses being identified as a function of the anchors of the redundant object hypotheses, using a target function assigned to the region proposal network; and fusing the selected object hypothesis with the identified redundant object hypotheses for the object detection.