Stereo Camera U-Map Labeling for Wall Vehicle Separation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing systems for detecting three-dimensional objects, such as vehicles and side walls, often mistakenly identify a side wall along a curved road as a vehicle or incorrectly detect overlapping objects as a single entity, leading to inaccurate object recognition.
Innovation Solution
The system employs a stereo camera with a U-map generating unit that creates high-resolution and low-resolution U-maps to accurately detect three-dimensional stretching objects by prioritizing depth direction continuity and using labeling processes to differentiate between walls and vehicles, even in overlapping scenarios.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a stereo camera device is used to detect objects, then the detection speed and coverage are improved, but false detection occurs where side walls are mistakenly identified as vehicles or pedestrians
Solution Approach 1:
The detection process is segmented into multiple stages: initial object detection, classification into stretching objects vs. non-stretching objects, and separate processing paths. This segmentation allows the system to handle different object types with appropriate methods, reducing false detection while maintaining high detection speed.
Solution Approach 2:
The patent transforms the two-dimensional image data into three-dimensional information by detecting stretching characteristics in the depth direction. By analyzing whether objects stretch across multiple depth positions, the system can distinguish between side walls (which stretch) and vehicles/pedestrians (which don't stretch), thereby improving detection accuracy without sacrificing speed.
2Measurement precision
If stretching objects are detected in the depth direction to distinguish walls from vehicles, then detection accuracy is improved, but the processing complexity increases
Solution Approach 1:
The processing complexity is managed by segmenting the detection pipeline: first detect all objects, then classify them by stretching characteristics, and finally apply different processing. This avoids the need for complex processing of all objects uniformly, reducing overall system complexity while maintaining high accuracy for stretching objects.
Solution Approach 2:
The system applies the stretching detection method selectively only to objects that exhibit stretching characteristics in the depth direction, rather than applying complex processing to all detected objects. This partial application of the complex method reduces processing complexity while maintaining detection accuracy for the critical case of distinguishing walls from vehicles.
3Device complexity
If overlapping objects are detected as single entities, then the detection process is simplified, but object recognition accuracy deteriorates
Solution Approach 1:
The system segments overlapping objects by detecting their individual stretching characteristics in the depth direction. Even when objects overlap in the two-dimensional image, their distinct depth positions and stretching patterns allow the system to separate and identify them individually, maintaining recognition accuracy without significantly complicating the detection process.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
A generating unit generates two-dimensional distribution information indicating a two-dimensional distribution of an object, in which an actual distance in a horizontal direction, a distance value indicating a distance in a depth direction, and a frequency value indicating an occurrence frequency of a distance value are related by using for example a disparity image. A labeling unit conducts a labeling process by conducting search on the two-dimensional distribution information multiple times, detecting a pixel having the frequency value that is more than a predetermined threshold during each search, and assigning a label that corresponds to the number of times of search. During this labeling process, a labeling process is conducted by selecting a pixel having a frequency value more than the above-described threshold from pixels that abut a pixel being searched in a search direction. A labeling process is selectively performed on pixels, whereby a labeling process may be conducted in accordance with the shape of an object. Objects are detected from two-dimensional distribution information by regarding pixels with the identical label assigned thereto as pixels of an identical object so that objects may be detected in a separate manner with accuracy.