Self-Checkout Scan-Omission Detection Using Motion and Skeleton Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In self-scanning systems for retail stores, detecting scan omissions due to fraudulent acts like shoplifting is challenging, especially when scans can occur at various locations and not all customers follow the scanning process, leading to increased processing loads and inaccurate fraud detection.
Innovation Solution
A fraud detection system using a fraud detection apparatus connected to cameras and user terminals, employing object and skeleton detection algorithms to count product placement and registration motions, and evaluate discrepancies to identify scan omissions, with alerts sent to store clerks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional deep learning training methods are used, then model accuracy is improved, but calculation efficiency deteriorates due to wasteful calculations and inability to switch data types
Solution Approach 1:
The patent applies dynamics by making the data type flexible and switchable during training. The system dynamically switches between FP32 and 8-bit data types based on the training phase, allowing the calculation precision to adapt to different training requirements rather than using a fixed data type throughout the entire training process.
Solution Approach 2:
The patent implements parameter changes by modifying the data type parameter during training. The system changes the data type from FP32 to 8-bit format at appropriate training stages, and can switch back when needed. This parameter change enables the system to achieve both high detection accuracy and improved calculation efficiency.
2Reliability
If training continues until failure is detected, then model robustness is improved, but training time and computational resources are wasted
Solution Approach 1:
The patent applies preliminary action by performing early detection of training failure through monitoring validation loss. Instead of continuing training until obvious failure symptoms appear, the system proactively detects when training is no longer improving and stops early, preventing wasteful continuation of ineffective training iterations.
Solution Approach 2:
The system implements feedback mechanisms by continuously monitoring validation loss during training. When the validation loss stops decreasing or begins to increase, the feedback signal triggers early stopping, allowing the system to adjust training duration based on actual performance feedback rather than fixed predetermined criteria.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An information processing program causes a computer to execute a process including: specifying, from an image that is captured by a camera, a person and a plurality of objects, generating, by inputting the image of the person into a machine learning model, skeleton information on the person, identifying, based on the plurality of objects and the skeleton information, a first feature value associated with one or more first motions of the person who retrieves an object from among the plurality of objects, identifying a second feature value associated with one or more objects registered to a first terminal by the person from among the plurality of object, and generating, based on a difference between the first feature value and the second feature value, an alert indicates that an object retrieved by the person is not registered in the first terminal.