Dynamic GPU Throttling for ATM Crowd Object Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
ATMs positioned in public areas face increased risks of theft and vandalism, and existing surveillance systems consume excessive power, making it difficult to identify perpetrators amidst crowds while maintaining efficient power usage.
Innovation Solution
A surveillance system for ATMs that uses an image sensor and machine learning models to dynamically activate a GPU for image recognition when needed, switching to a CPU for lower power consumption when fewer objects are detected, thereby optimizing power usage and identifying potential threats.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a surveillance system with image recognition is continuously activated to detect and identify potential perpetrators, then security monitoring capability is improved, but power consumption increases excessively
Solution Approach 1:
The system dynamically switches between CPU and GPU based on real-time monitoring needs. The GPU is activated only when a crowd is detected by the image sensor, and deactivated when no crowd is present, making the system's computational power dynamic rather than static.
Solution Approach 2:
The surveillance system operates in periodic cycles: the image sensor continuously captures images, the CPU processes these images to detect crowds, and only when a crowd is detected does the system activate the GPU for intensive object recognition. This periodic activation pattern reduces overall power consumption while maintaining security.
2Measurement precision
If a GPU is used for real-time object recognition in crowded areas, then identification accuracy is improved, but power consumption increases
Solution Approach 1:
The system applies different processing qualities to different scenarios: when no crowd is detected, basic CPU processing suffices; when a crowd is detected in the monitored area, the system locally activates GPU resources specifically for that period to achieve high-precision object recognition only where and when needed.
Solution Approach 2:
Instead of continuously using the full GPU power for all frames, the system applies partial action by activating the GPU only for frames containing crowds. This partial activation provides sufficient identification accuracy for security purposes while avoiding the excessive power consumption of continuous GPU operation.
3Use of energy by moving object
If the CPU processes all video frames continuously, then power consumption is reduced, but object recognition speed and accuracy decrease
Solution Approach 1:
The system dynamically adjusts processing resources based on scene complexity. For simple frames without crowds, the CPU handles processing efficiently with low power consumption. When crowds are detected, the system dynamically transitions to GPU-accelerated processing to maintain high recognition speed and accuracy.
Solution Approach 2:
The CPU performs preliminary action by continuously analyzing frames to detect the presence of crowds before triggering the more resource-intensive GPU processing. This preliminary detection step allows the system to prepare for high-speed recognition only when necessary, optimizing the balance between power consumption and recognition performance.
Data Source
AI summary
Apparatus and methods for dynamically controlling application of a graphic processing unit (“GPU”) based on the current processing needs is described. A number of objects in each image may be detected. A machine learning model may be utilized to detect and classify objects in each image. Based on the number of detected objects, the GPU could be switched on to provide increased image processing performance. Dynamically turning the GPU on/off may reduce power consumption. When the GPU is turned off, a central processing unit (“CPU”) may be used to perform image processing.