Hybrid Activity Recognition System for Video Surveillance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing activity recognition systems face challenges in accurately recognizing human activities in real-time video surveillance due to the limitations of traditional algorithms, which are often resource-intensive and prone to network delays when relying solely on cloud computation, and lack effective integration between local devices and cloud servers.
Innovation Solution
A hybrid system that utilizes a local device for initial human detection, re-identification, and stranger detection, while transmitting relevant video portions to a cloud server for complex activity recognition, leveraging advanced algorithms like the Hierarchy Pose Temporal Model for accurate activity classification and alarm triggering.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If complex recognition algorithms are executed in a Cloud server, then recognition accuracy is improved, but network delay and bandwidth consumption increase
Solution Approach 1:
The patent divides the activity recognition system into two segments: local device execution of simple detection algorithms (human detection, re-identification, stranger detection) and cloud server execution of complex recognition algorithms. This segmentation allows basic filtering to occur locally without network delay, while complex analysis benefits from cloud computing power, thus resolving the contradiction between recognition accuracy and network delay
Solution Approach 2:
The local device performs preliminary detection and filtering actions (detecting human presence, determining if the same person is present, identifying strangers) before transmitting video to the cloud. This preliminary action reduces the amount of data requiring cloud processing and minimizes network delay for critical detection functions
2Adaptability or versatility
If complex recognition algorithms are executed in a Cloud server, then more human activities can be recognized, but network bandwidth consumption increases significantly
Solution Approach 1:
The patent extracts and removes basic detection functions (human detection, re-identification, stranger detection) from the cloud server and places them in the local device. This extraction reduces the bandwidth consumption required for cloud processing while maintaining comprehensive activity recognition capabilities through cloud-based complex algorithms
Solution Approach 2:
The system segments processing tasks by complexity: simple detection tasks are handled locally to conserve bandwidth, while complex activity recognition tasks are handled in the cloud to maintain versatility. This segmentation resolves the contradiction between comprehensive activity recognition and bandwidth consumption
3Loss of energy
If traditional algorithms are executed in local devices, then network bandwidth is reduced, but recognition accuracy decreases
Solution Approach 1:
The patent segments the algorithm portfolio into simple detection algorithms executed locally (reducing bandwidth) and complex recognition algorithms executed in the cloud (maintaining accuracy). This segmentation resolves the contradiction between bandwidth conservation and recognition accuracy by strategically placing different algorithm types at different locations
Data Source
AI summary
An activity recognition system may comprise a local device and a server. The local device may be configured to: obtain a video; determine whether at least one human is present in one or more frames of the video; in response to determining the at least one human present in the video, determine if the at least one human in each of the frames corresponds to the same person; in response to determining that the at least one human in each of the frames corresponds to the same person, determine if the person is a stranger; and in response to determining that the person is not a stranger, transmit at least a portion of the video to the server. The server may be configured to: predict an activity class associated with the person; and trigger an alarm based on the activity class.