Object Detection via Part Area Probability Distributions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection techniques, such as the sliding window approach and ensemble learning methods, face challenges in accurately determining the position of objects, especially when objects are partially hidden or when the window size differs significantly from the object size, leading to false detections and decreased accuracy.
Innovation Solution
An object detection apparatus and method that generates appearance and absence probability distributions for part areas based on their frequency of appearance in images, allowing for accurate determination of object areas by considering the positional relationships and penalties for false detections.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If the sliding window approach is used to detect objects, then the detection process can be implemented with a simple rectangular window, but it becomes difficult to accurately determine object position when the object is partially hidden or when the window size is largely different from the object size
Solution Approach 1:
The patent segments the object detection problem into multiple part areas (e.g., head, body, limbs for human detection). Each part area is detected and evaluated independently, then integrated to form the complete object detection result. This segmentation allows accurate detection even when parts are partially hidden, as other visible parts can still contribute to the overall detection confidence.
Solution Approach 2:
The patent introduces a new dimension of evaluation by considering the positional relationships between multiple part areas. Instead of evaluating a single rectangular window, the system evaluates multiple part areas in different spatial positions and their relative configurations, adding spatial relationship dimension to the detection process.
2Reliability
If ensemble learning with multiple decision trees is used to detect parts, then detection robustness is improved, but false detections occur when positional relationships between parts are not considered
Solution Approach 1:
The patent implements feedback mechanisms where the detection results of individual part areas are evaluated not only by their own confidence scores but also by their positional relationships with other detected parts. The system provides feedback to adjust the overall object detection decision based on whether the detected parts form a coherent spatial configuration consistent with the target object structure.
Solution Approach 2:
The patent performs preliminary detection of multiple part areas independently before integrating them into a complete object detection. This preliminary action allows each part to be evaluated separately with robust ensemble learning, and then the results are combined with spatial relationship constraints to eliminate false detections that don't satisfy the expected positional configurations.
3Measurement precision
If the constellation model is used to evaluate positional relationships between parts, then object detection accuracy is improved, but the detection area may become larger than the actual object area
Solution Approach 1:
The patent applies local quality evaluation by assigning different weights and evaluation criteria to different part areas based on their individual characteristics and reliability. Instead of treating all parts uniformly, the system evaluates each part area with appropriate local constraints and combines them to form the final object detection, ensuring the detected area closely matches the actual object boundaries.
Data Source
AI summary
An object detection apparatus, etc., capable of detecting an object area with greater precision is disclosed. Such an object detection apparatus is provided with: a part area indication means for indicating a part area which is an area including a target part among parts forming an object including an detection-target object, from a plurality of images including the object; an appearance probability distribution generation means for generating an appearance probability distribution and the absence probability distribution of the part area based on the appearance frequency of the part area associated with each position in the images; and an object determination means for determining, in an input image, the area including the object, with reference to the appearance probability distribution and the absence probability distribution of the part area.