Distributed Camera Classification to Cut Video Bandwidth
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Surveillance systems face challenges in efficiently searching for and locating items of interest due to the high communication bandwidth and computing power requirements for processing large amounts of video data from security cameras.
Innovation Solution
A distributed camera system is implemented, where each camera is equipped with a Deep Learning Accelerator (DLA) to process video images using an item model, reducing the need for centralized computing and communication bandwidth by performing local classification and transmission only of relevant data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If centralized processing is used to analyze video data from security cameras, then comprehensive monitoring coverage is achieved, but communication bandwidth and computing power requirements become excessively high
Solution Approach 1:
The system segments the centralized processing function by deploying Deep Learning Accelerators at each camera node. Each camera independently processes its own video data locally using the DLA, dividing the overall computational task into distributed units. This segmentation eliminates the need to transmit all video data to a central server, thereby reducing communication bandwidth requirements while maintaining comprehensive monitoring coverage.
Solution Approach 2:
Each camera is equipped with a Deep Learning Accelerator that enables it to autonomously process and classify video data locally. The camera performs self-service by executing classification models on its own captured images, determining whether to transmit data based on local analysis results. This self-service capability reduces dependency on centralized computing resources and minimizes communication overhead.
2Extent of automation
If all video data is transmitted to central station for processing, then centralized analysis capability is maintained, but data transmission volume and time delay increase
Solution Approach 1:
The system performs preliminary classification action at the camera level before data transmission. Each camera uses its embedded DLA to pre-process and classify video frames locally, filtering out irrelevant data beforehand. Only data that passes the local classification threshold is transmitted to the central station, significantly reducing transmission volume and time while preserving the ability for centralized analysis of confirmed objects of interest.
3Productivity
If deep learning models are deployed at each camera, then local processing efficiency improves, but device complexity at camera nodes increases
Solution Approach 1:
Instead of implementing complex full-depth neural networks at each camera, the system uses copies of pre-trained Deep Learning Accelerator hardware modules. These DLA copies are specialized chips that can be programmed with different classification models as needed. The copying approach maintains high processing efficiency while keeping the hardware design modular and manageable, as the same DLA architecture can be reused across all camera nodes with different software configurations.
Data Source
AI summary
A system having a central station and a plurality of cameras installed various locations. To search for and locate an item of interest, the central station generates and sends an item model to the cameras. When stored in a camera, the item model causes a logic circuit of the camera (e.g., a deep learning accelerator) to use image data, received from an image sensor for storing in a memory device of the camera, as an input to an artificial neural network. The logic circuit performs the matrix computation of the artificial neural network to generate a classification of whether the images are relevant to the item of interest characterized by the item model. If so, the camera transmits the relevant images to the central station for further processing to determine a real time location of the item of interest.