Camera Object Retrieval Using Segmented Discriminators
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for retrieving a specific person from large-scale monitoring camera videos are inefficient, particularly in detecting the target person while avoiding unnecessary detections and handling varying appearances and backgrounds, leading to complex discriminator arrangements and detection errors.
Innovation Solution
An image retrieval apparatus and method that sequentially selects cameras, acquires background and query images, learns a discriminator using these images, and retrieves the target object using the learned discriminator, efficiently narrowing down image patterns without overcomplicating the discriminator arrangement.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If human region detection is performed on all images to retrieve a specific person, then the retrieval can be performed, but unnecessary persons are detected and the discriminator arrangement becomes overly complex
Solution Approach 1:
The detection process is segmented into two distinct stages: first, a simple discriminator performs rough detection to identify potential human regions, then a second, more complex discriminator performs precise detection only on those candidate regions. This segmentation allows the complex discriminator to focus only on relevant areas rather than processing entire images, reducing overall system complexity while maintaining high retrieval accuracy.
Solution Approach 2:
The simple discriminator performs preliminary detection to identify candidate human regions before the complex discriminator is applied. This preliminary action filters out non-human regions and backgrounds, so that the complex discriminator only needs to process a small subset of candidate regions, thereby simplifying the overall discriminator arrangement while preserving detection precision.
2Device complexity
If a simple discriminator is used for detection, then the arrangement is simple, but detection errors occur in background portions
Solution Approach 1:
The detection system is divided into two discriminator stages with different complexity levels. The simple discriminator handles the majority of background rejection, while the second, more accurate discriminator corrects errors in candidate regions. This segmentation allows each discriminator to be optimized for its specific function, maintaining overall system simplicity while improving reliability.
Solution Approach 2:
The simple discriminator acts as an intermediary that filters out obvious non-human regions before the final detection is made. It provides a preliminary classification that reduces the workload for the more accurate discriminator, allowing the system to maintain simplicity while achieving high detection accuracy through the intermediary filtering stage.
3Reliability
If detection is performed on all images to ensure no target is missed, then detection completeness is improved, but the retrieval time increases
Solution Approach 1:
The simple discriminator performs preliminary filtering to quickly identify candidate human regions, eliminating the need for complex processing on background portions. This preliminary action maintains detection completeness by ensuring all potential targets are identified, while significantly reducing retrieval time by avoiding unnecessary processing of non-human regions.
Solution Approach 2:
The system performs detection in two passes: first a quick partial detection using the simple discriminator to identify candidate regions, then a more thorough detection using the complex discriminator only on those candidates. This partial action approach ensures no targets are missed while minimizing the time spent on regions that definitely don't contain targets.
Data Source
AI summary
Each of a plurality of cameras is sequentially selected. An image that does not include a retrieval target object is acquired, as a background image, from an image group captured by a selected camera. An image including the retrieval target object is acquired as a query image. A discriminator being used for identifying the retrieval target object from the image captured by the selected camera is learned using the background image and the query image. The retrieval target object is retrieved from the image group captured by the selected camera using the learned discriminator.