Monocular Floor Region Detection Using Object Footprint Geometric Cues
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting floor regions in images captured by a static camera are limited in handling partially occluded targets and complex scenes, as they fail to accurately identify horizontal occlusion boundaries parallel to target motion and do not consider disconnected floor segments or 'sittable' and 'walkable' surfaces.
Innovation Solution
A method that detects objects in an image, determines regions where the lower part of the object exists, calculates the confidence of these regions as potential supporting surface regions, and selects regions based on their similarity to detect the supporting surface, using geometric cues from moving targets without requiring multiple viewpoints or prior knowledge of camera geometry.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If existing floor detection methods use simple footprint-based approaches, then detection is straightforward in simple scenes, but accuracy deteriorates in complex scenes with occlusions and disconnected floor segments
Solution Approach 1:
The patent segments the floor detection problem into multiple components: identifying seed pixels from footprints, performing region growing from seeds, detecting occlusion boundaries using persistent edges, and handling disconnected floor segments separately. This segmentation allows the system to address each aspect independently, improving overall accuracy in complex scenes while maintaining reasonable complexity.
Solution Approach 2:
The patent transitions from 2D footprint analysis to 3D geometric reasoning by inferring occlusion boundaries and floor planes in three-dimensional space. By using depth information and spatial relationships, the system can distinguish between foreground objects and floor regions even when they overlap in the 2D image plane, significantly improving detection accuracy in complex scenes.
2Use of energy by moving object
If footprint-based methods are used for floor detection, then the approach is computationally simple, but it fails to handle partially occluded targets and disconnected floor segments
Solution Approach 1:
The patent introduces intermediate structures including depth maps, occlusion boundary detectors, and floor plane models that act as mediators between the simple footprint detection and the final floor region identification. These intermediaries enable the system to handle occlusions and disconnected segments without requiring complex computational approaches throughout the entire pipeline.
Solution Approach 2:
The patent changes key parameters including using confidence thresholds for seed selection, adjusting region growing criteria based on local geometry, and modifying detection sensitivity for different scene conditions. These parameter adjustments allow the system to maintain computational efficiency while adapting to various occlusion scenarios and floor configurations.
3Measurement precision
If multiple viewpoints or prior camera geometry knowledge are required, then detection accuracy may improve, but system complexity and requirements increase
Solution Approach 1:
The patent enables the monocular camera system to determine its own geometric properties by analyzing the scene content itself. The system infers floor planes, occlusion boundaries, and spatial relationships directly from the single viewpoint image, without requiring external calibration data or multiple cameras. This self-service approach maintains accuracy while minimizing system complexity.
Solution Approach 2:
The patent creates a universal floor detection algorithm that works across diverse scene types (indoor/outdoor, various lighting conditions, different floor materials) using a single camera viewpoint. The method universally handles occlusions, disconnected segments, and varying geometries through geometric reasoning that adapts to each scene without requiring scene-specific calibration or multiple specialized sensors.
Data Source
AI summary
A method of detecting a supporting surface region in an image captured by a camera is disclosed. An object in the image is detected. One or more regions of the image in which a lower part of the detected object exists are determined. A degree of confidence for each of the regions is determined. The degree of confidence indicates likelihood of a corresponding region being a supporting surface region. One or more of the regions are selected based on each corresponding degree of confidence. Similarity of other regions in the image to at least one of the selected regions is determined. The supporting surface region is detected based on the determined similarity.