Monitoring Device for Left Object Detection via Feature Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection methods, such as background difference and histogram of oriented gradients (HoG) with machine learning, struggle to accurately distinguish stationary objects from noise and environmental changes, leading to false alarms and reduced detection accuracy, especially when sunlight conditions change or when objects are partially occluded.
Innovation Solution
A monitoring device that generates a reference image, extracts foreground regions, and uses feature accumulation and clustering to identify stationary objects, determining whether they are left objects based on feature analysis and environmental evaluation, with a management unit to issue alarms only when necessary.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If background difference method is used for object detection, then detection speed is improved, but detection accuracy deteriorates due to false alarms from environmental changes
Solution Approach 1:
The patent segments the detection process into multiple stages: initial background difference detection to identify candidate regions, followed by HoG feature extraction and machine learning classification to verify actual objects. This multi-stage segmentation allows fast initial screening while maintaining high accuracy through subsequent verification steps.
Solution Approach 2:
The patent introduces HoG feature extraction and machine learning classification as intermediary steps between background difference detection and final object identification. These intermediaries filter out false alarms from environmental changes while preserving true objects, resolving the contradiction between speed and accuracy.
2Measurement precision
If HoG + machine learning is used to identify left objects, then detection accuracy is improved, but device complexity increases due to learning data requirements
Solution Approach 1:
The patent applies machine learning selectively only to candidate regions identified by background difference detection, rather than processing the entire image. This partial application reduces computational complexity while maintaining high detection accuracy for left objects.
Solution Approach 2:
The patent performs preliminary background difference detection to identify candidate regions before applying the more complex HoG + machine learning process. This preliminary action reduces the data volume requiring complex processing, thereby reducing overall system complexity while maintaining accuracy.
3Reliability
If template matching is used to handle occluded objects, then robustness is improved, but false alarms increase due to wrong timing registration
Solution Approach 1:
The patent uses feedback from the machine learning classification results to adjust template matching parameters and timing. When the system detects potential wrong timing registration through classification confidence scores, it adjusts the template registration strategy, reducing false alarms while maintaining robustness against occlusion.
Solution Approach 2:
The patent dynamically adjusts the template matching process based on real-time detection conditions. When occlusion is detected, the system adapts template registration timing and parameters dynamically rather than using fixed rules, thereby reducing false alarms caused by wrong timing while maintaining robustness.
Data Source
AI summary
An image monitoring system includes: a reference image generation unit generating a reference image serving as a reference image, based on an input image; a foreground extraction unit detecting, from the input image, a foreground region different from the reference image; a stationary region extraction unit extracting a stationary region from the foreground region; a feature quantity extraction unit extracting a feature quantity of the stationary region; a feature quantity recording unit recording feature quantities as a function of time; a stationary object detection unit clustering pixels, recorded feature quantities of which meet predetermined criteria, and detecting a stationary object; a left object determination unit determining whether the stationary object is a left object, based on a feature quantity of the stationary object and evaluation of the surrounding environment; and a left object management unit managing attributes of left objects, and issuing a report when a certain condition is met.