Foreground Detection Using Time-Decaying Likelihood
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image processing methods fail to accurately distinguish between a static object and the background, leading to unwanted detection of a person as background if they remain still for too long, while static objects like abandoned items are continuously detected as foreground, disrupting the ability to detect specific subjects like people.
Innovation Solution
An information processing apparatus and method that uses likelihood information to determine the probability of a region representing a subject, updating background-model information to differentiate between subjects and static objects, ensuring continuous detection of people while correctly identifying static objects as background after a certain time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If background model is updated at given intervals to adapt to background changes, then background adaptability is improved, but static objects are continuously detected as foreground indefinitely
Solution Approach 1:
The patent introduces a likelihood parameter that changes over time for each region. When a region is first detected as foreground, its likelihood is set to a high value. In subsequent frames, the likelihood is gradually decreased based on the frame number. This parameter change allows the system to initially detect new objects (including static ones) as foreground while eventually classifying them as background, resolving the contradiction between adapting to background changes and reliably detecting subjects.
2Measurement precision
If region with intrusion object is excluded from background-update target to prevent intrusion object from being considered as background, then intrusion detection accuracy is improved, but static objects left behind are continuously detected as foreground
Solution Approach 1:
The patent makes the background model dynamic by introducing time-dependent likelihood values for each region. Instead of statically excluding regions with intrusion objects from background updates, the system allows all regions to have their likelihood values dynamically adjusted over time. This enables the system to adapt to background changes while automatically transitioning static objects from foreground to background classification, eliminating the need for manual region exclusion and resolving the time loss issue.
3Measurement precision
If time information is added as object-determination condition to distinguish temporary objects from background, then temporary object detection is improved, but persons stopping longer than predetermined time are unwantedly considered as background
Solution Approach 1:
The patent applies local quality by assigning different likelihood decay characteristics to different regions based on their initial detection as foreground. Regions with high initial likelihood (indicating potential subjects) maintain higher likelihood values longer, while regions with lower initial likelihood (indicating background) have their likelihood values decreased more rapidly. This localized differentiation allows the system to distinguish between temporary objects and persons, preventing persons from being misclassified as background while still detecting temporary objects accurately.
Data Source
AI summary
There is provided an information processing apparatus for detecting a foreground portion from an acquired image. Feature-amount information of a past acquired image and likelihood information indicating a probability that a subject is represented is stored as background-model information. A foreground region is detected based on the likelihood information and a result of comparing the feature amount extracted from the acquired image with the feature-amount information contained in the background-model information. For each of the detected foreground region, a likelihood indicating a probability that the region represents the subject is calculated based on information prepared in advance related to the subject.