Moving Object Detection via Epipolar Direction Averaging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting moving objects from images captured by a moving camera struggle with accurately distinguishing between object motion and parallax motion, especially for small object features, leading to unreliable elimination of false detections.
Innovation Solution
A computer-implemented method that estimates the direction of epipolar movement by averaging flow directions from a sequence of images, using difference images to reduce noise and identify regions with pixel values exceeding a threshold, and filters out regions matching along this direction to retain only those due to object movement.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If motion vectors are estimated from images to eliminate parallax, then false moving object detections can be eliminated, but for small object features the elimination becomes unreliable
Solution Approach 1:
The patent introduces an intermediary approach by using difference images as a mediator between the original images and the motion estimation process. By computing difference images first (which highlight only the changing regions), the subsequent motion vector estimation is performed on a simplified, noise-reduced representation. This intermediary step allows for more reliable detection of small object features while maintaining robust parallax elimination, as the difference images contain only the essential motion information without the full complexity of the original image content.
2Reliability
If image regions are matched to determine epipolar movement, then parallax can be eliminated, but the detection becomes unstable and cannot reliably match small regions
Solution Approach 1:
The patent applies segmentation by dividing the image processing into distinct stages: first computing difference images to segment out the moving regions, then performing motion estimation only on these segmented difference regions. This segmentation allows small object features to be isolated and processed independently, improving both the stability of detection and the ability to match small regions. The difference image computation effectively segments the problem space, allowing reliable processing of small features that would be lost in full-image processing.
3Measurement precision
If motion vectors are computed from original images, then comprehensive motion information is obtained, but noise and outliers reduce the accuracy of epipolar direction estimation
Solution Approach 1:
The patent extracts only the essential motion information by computing difference images that isolate changing regions from the full image data. This extraction process removes the harmful noise and outliers present in the original images, retaining only the signal related to actual motion. By taking out only the relevant motion components through difference image computation, the subsequent epipolar direction estimation is performed on clean, noise-free data, significantly improving measurement precision while filtering out object-affected harmful factors.
Data Source
Figure 1
Figure 2
Figure 3~3b
AI summary
Moving objects are detected from a temporal sequence of images captured from successive viewpoints. Image flow vectors are computed from the temporal sequence of images, for example of flow between difference images derived from the sequence. An average direction of the image flow vectors associated with displacement of image areas between a first image and a second image derived from temporal sequence of images is computed. Preferably, a method of averaging is used that is designed to suppress the effect of outliers. Regions are determined in the first image. For each of the identified regions along range of positions along the computed average direction from a position of the identified region, it is determined whether or not there is a match between first image content in the identified region in the first image and second image content in a patch located within said range of positions in the second image. The set of identified regions is reduced by removing identified regions for which it was determined in said determining step that there was a match.