Multi-Camera Person Identification Using Feature Graphs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing person tracking techniques in stores face inaccuracies due to varying image characteristics among different monitoring cameras, leading to erroneous identification of individuals, which is exacerbated by differences in camera angles, brightness, clothing, and store backgrounds, making it impractical to train models for each store individually.
Innovation Solution
A method that utilizes machine learning to generate training data by leveraging overlapping image capturing regions of multiple cameras, correcting person features using pseudo labels, and optimizing the selection of samples to maximize similar samples, thereby improving the accuracy of person identification models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a person identification model is trained using public datasets with generic image characteristics, then the model can be constructed quickly and with less cost, but the inference accuracy decreases when applied to stores with different image characteristics such as varying camera angles, brightness, backgrounds, and customer demographics
Solution Approach 1:
The patent applies parameter changes by transforming the image characteristics of the public dataset to match the target store's environment. Specifically, it adjusts brightness, contrast, saturation, and other visual parameters of the training images to align with the actual store conditions including camera angles, lighting conditions, and background characteristics. This allows the model trained on generic data to adapt to store-specific conditions without requiring complete retraining from scratch.
Solution Approach 2:
The patent performs preliminary actions by pre-processing the public dataset to extract and generate synthetic store-specific training data before model training. It creates augmented training samples that simulate the target store environment in advance, including generating synthetic backgrounds, adjusting image characteristics, and preparing feature vectors that reflect the actual deployment conditions. This preliminary preparation ensures the model is better adapted when deployed.
2Measurement precision
If training data is collected from actual store environments for each specific store, then the person identification accuracy improves, but the cost and time required for data collection and model training becomes prohibitively high
Solution Approach 1:
The patent uses copying by creating synthetic copies of training data that simulate actual store conditions. Instead of collecting real images from each store, it generates synthetic training samples by transforming public dataset images to replicate store environments including camera perspectives, lighting conditions, and background characteristics. These synthetic copies serve as substitutes for actual store data, reducing collection costs while maintaining training effectiveness.
Solution Approach 2:
The patent introduces an intermediary processing layer that transforms generic public dataset images into store-specific training data. This intermediary process includes image augmentation, characteristic adjustment, and synthetic background generation that bridges the gap between public datasets and store-specific requirements. The intermediary transformation enables reuse of public data while adapting it to specific deployment environments.
3Adaptability or versatility
If multiple models are trained for different store characteristics to cover all variations, then the identification accuracy across diverse stores improves, but the device complexity and computational resources required become unmanageable
Solution Approach 1:
The patent achieves universality by creating a single person identification model that can adapt to multiple store characteristics through parameter transformation. Instead of maintaining separate models for different stores, the system uses a universal model with adjustable parameters that can be configured to match various store environments including different camera angles, lighting conditions, and demographics. This single model serves multiple functions across diverse deployment scenarios.
Solution Approach 2:
The patent applies dynamics by making the model parameters adaptive rather than fixed. The system dynamically adjusts image characteristics and model parameters based on the target store environment, allowing the same model to flexibly adapt to varying conditions. This dynamic parameter adjustment replaces the need for multiple static models, reducing complexity while maintaining adaptability across different stores.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A determination program that causes at least one computer to execute a process, the process includes obtaining a plurality of pair images of a person obtained from a overlapping region of images captured by each of a plurality of cameras; generating a directed graph including nodes corresponding to person features obtained from each of a plurality of person images included in the plurality of obtained pair images; acquiring weights of links between the nodes in the generated directed graph based on a number of person images with similar person features between the nodes; and determining a combination of the person features in which a number of the person images with the similar person features in the plurality of pair images is maximized based on the acquired weights of the links.