Object Detection Region Selection for Constant-Speed Edge Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection methods struggle to maintain high accuracy and processing speed when applied to high-definition images in resource-limited environments, such as edge terminals, due to increased calculation demands and errors in image synthesis or division.
Innovation Solution
An object detection device and method that adaptively selects a fixed number of rectangles for object detection based on object distribution and past frames, using density estimation and priority scoring to reduce the number of rectangles processed, while maintaining a constant processing speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If image division is applied to high-definition video for object detection, then detection coverage is improved, but processing time increases and accuracy decreases due to large number of divisions and synthesis errors
Solution Approach 1:
The patent applies segmentation by dividing the image into multiple regions and performing object detection on each region separately. This allows parallel processing of different image portions, reducing overall processing time while maintaining comprehensive detection coverage across the entire high-definition image.
Solution Approach 2:
The patent performs preliminary processing by generating candidate bounding boxes and filtering them before final object detection. This preliminary action reduces the number of regions requiring full detection processing, thereby decreasing processing time while preserving detection accuracy for relevant objects.
2Productivity
If adaptive image division is used to reduce division number, then processing speed improves, but detection accuracy may decline when division number becomes too large
Solution Approach 1:
The patent dynamically adjusts the number and size of image divisions based on the detected objects and scene characteristics. This dynamic adaptation allows the system to process fewer regions when objects are concentrated, improving processing speed, while maintaining sufficient division granularity when needed to preserve detection accuracy.
Solution Approach 2:
The patent changes processing parameters such as the number of divisions and region sizes based on input image characteristics and object distribution. This parameter adjustment optimizes the balance between processing speed and detection accuracy for different scenarios.
3Reliability
If object detection is applied to all divided images, then detection completeness is improved, but calculation resources are exhausted in resource-limited environments
Solution Approach 1:
The patent extracts and processes only the most relevant image regions containing potential objects, rather than uniformly processing all divided images. This extraction approach reduces calculation resource consumption by focusing computational efforts on areas with high detection probability while maintaining detection completeness for actual objects.
Solution Approach 2:
The patent applies partial processing by performing full object detection only on regions with high object probability, while using simplified or skipped processing for low-probability regions. This partial action reduces overall calculation resource consumption while maintaining sufficient detection completeness for actual objects.
Data Source
AI summary
Object detection with high accuracy can be realized while maintaining a constant processing speed even in a limited environment of resources.An object detection device includes a rectangle extraction unit that extracts a plurality of rectangles to be candidates to which object detection is applied from an input image, a rectangle selection unit that selects a fixed number of rectangles to which the object detection is applied from among the rectangle candidates extracted from the rectangle extraction unit, and an object detection unit that performs the object detection on the rectangle selected by the rectangle selection unit to output metadata including at least a class, reliability, and a bounding box of the object included in the input image as an object detection result.