Action Recognition Device Using Target Queue for GPU Load Management
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional action recognition techniques face performance issues as the number of action analysis targets increases, leading to increased GPU usage in embedded devices, affecting overall system efficiency.
Innovation Solution
An action recognition device and method utilizing a separate edge device for asynchronous action recognition, which includes a camera, a target queue, and a computing device that updates and manages targets based on object tracking results, requesting action reasoning from a reasoner when conditions are met, and outputs recognition results while managing resources to avoid duplicate reasoning and maintain performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If the number of action analysis targets increases, then the coverage of action recognition improves, but the GPU usage increases and system performance deteriorates
Solution Approach 1:
The system segments the action recognition process by introducing a target queue that divides targets into different processing states (waiting, being processed, completed). This segmentation allows the reasoner to process targets asynchronously rather than handling all targets simultaneously, thereby maintaining system performance while increasing the number of targets that can be analyzed.
Solution Approach 2:
The system implements dynamic target management where the target queue adapts its content based on processing status. Targets are dynamically added to, removed from, and transitioned between different processing stages. This dynamic approach allows the system to handle variable numbers of targets efficiently, preventing GPU overload while maintaining comprehensive coverage.
2Quantity of substance
If action recognition is performed on multiple objects simultaneously, then the comprehensiveness of monitoring improves, but the processing speed decreases due to resource constraints
Solution Approach 1:
The system implements periodic action processing where the reasoner processes targets in periodic batches rather than continuously. The target queue manages this periodic processing by holding targets until the reasoner is ready, allowing the system to maintain high processing speed while monitoring multiple objects through structured, periodic analysis cycles.
Solution Approach 2:
The target queue serves as an intermediary between the image capture device and the reasoner. It buffers and manages the flow of target data, allowing the reasoner to process objects at its own pace without being constrained by the rate at which new targets are generated. This intermediary mechanism decouples the production rate from the processing rate, maintaining speed while handling multiple objects.
3Loss of time
If the reasoner processes all targets immediately, then the responsiveness improves, but the resource consumption increases and causes system instability
Solution Approach 1:
The system performs preliminary actions by pre-processing and queueing targets before they reach the reasoner. The target queue prepares targets in advance, organizing them for efficient processing. This preliminary organization allows the reasoner to process targets with minimal delay when ready, maintaining responsiveness while avoiding resource exhaustion by not immediately overloading the GPU with all targets.
Solution Approach 2:
The target queue implements self-service by autonomously managing its own content and processing flow. It automatically adds new targets, removes processed targets, and manages the queue state without requiring constant external intervention. This self-managing approach allows the system to respond efficiently to new targets while the queue intelligently regulates the flow to prevent GPU overload and energy waste.
Data Source
AI summary
Disclosed are an action recognition device and an action recognition method for recognizing actions of multiple objects in real time. The action recognition device includes a camera for capturing an image, a target queue storing therein data about at least one target, a reasoner configured to perform action reasoning, and a computing device electrically connected to the camera, the target queue and the reasoner, wherein the computing device may update the target queue based on an object tracking result on the image, extract a target from the target queue, request the reasoner to perform action reasoning on the extracted target, and output an action recognition result based on an action reasoning result provided from the reasoner.