Polygonal Object Detection Window for Autonomous Driving Precision
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current object detection technologies in autonomous driving systems face challenges in accurately recognizing the position of tilted objects within images due to the need for larger rectangular windows, which include unnecessary areas, and struggle with real-time processing requirements for semantic segmentation.
Innovation Solution
An electronic device with a processor that identifies areas of change and stability in moving images, merges image information from stable areas to enhance object detection using polygonal windows with five or more angles, allowing for more accurate object positioning and distance measurement while minimizing computational load.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Area of stationary object
If rectangular windows are used to detect tilted objects, then the detection coverage is improved, but the detection precision deteriorates due to including unnecessary areas
Solution Approach 1:
The patent segments the detection window from a fixed rectangular shape to a dynamic polygonal shape that adapts to the detected object's orientation and boundaries. This segmentation allows the detection window to be divided into relevant and irrelevant regions, improving precision by excluding unnecessary areas while maintaining coverage of the tilted object.
Solution Approach 2:
The patent transitions from two-dimensional rectangular detection windows to polygonal windows with five or more vertices, adding geometric flexibility. This dimensional change in the window shape space enables precise fitting around tilted objects of various orientations, resolving the contradiction between coverage and precision.
2Measurement precision
If semantic segmentation is applied to accurately extract objects, then the detection precision is improved, but the processing speed deteriorates due to large computation amount
Solution Approach 1:
Instead of applying full semantic segmentation to the entire image, the patent applies object detection only to relevant regions identified by the polygonal window. This partial action approach maintains high precision by focusing computational resources on areas containing objects, while avoiding the excessive computation of segmenting the entire image, thus enabling real-time processing.
Solution Approach 2:
The patent performs preliminary object detection to identify potential object locations and orientations before applying more precise detection methods. This preliminary action using polygonal windows pre-segments the image into relevant and irrelevant regions, reducing the computational load for subsequent precise extraction while maintaining accuracy.
3Area of stationary object
If larger rectangular windows are used for tilted objects, then the detection coverage is improved, but the computational load increases due to processing larger areas
Solution Approach 1:
The patent segments the detection area into a polygonal shape that precisely follows the object's boundaries and orientation. This segmentation eliminates the need to process the entire large rectangular area, reducing computational energy consumption by focusing only on the relevant portion of the image while maintaining adequate coverage of the tilted object.
Solution Approach 2:
The patent applies different detection strategies to different regions of the image based on local characteristics. High-priority processing is applied only to regions containing detected objects using polygonal windows, while other regions receive minimal or no processing. This local quality approach reduces overall computational energy consumption while maintaining detection coverage where needed.
Data Source
AI summary
An object detection device is disclosed. The object detection device of the present disclosure comprises a memory and at least one processor operatively connected to the memory, wherein the at least one processor can acquire a video, identify, in the video, a region in which content is changed in real time and a region in which content is not changed in real time, acquire image information about the video, and merge the image information into the region in which content is not changed in real time.