Action Recognition Camera Server Distributed Processing

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Action recognition in video sequences requires excessive processing power, making it unsuitable for digital video cameras and resulting in high bandwidth usage when performed on servers, which necessitates an efficient method to distribute processing power and reduce bandwidth consumption.

Innovation Solution

A method that uses a combination of contextual/spatial action recognition algorithms on the camera for initial detection and temporal action recognition algorithms on a server, where the camera extracts and transfers only relevant action video sequences, allowing for distributed processing without continuous video streaming.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If action recognition is performed on the digital video camera, then processing power requirement increases, but the camera becomes unsuitable for the task due to excessive processing demands

Engineering Contradiction:
Improveaction recognition accuracyVSAvoidprocessing power requirement
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The action recognition process is segmented into two distinct stages: a first action recognition algorithm runs on the digital video camera to detect action candidates, while a second, more sophisticated algorithm runs on a server to verify and classify those candidates. This segmentation allows each component to operate within its processing capabilities while achieving high overall accuracy.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If action recognition is performed on a server, then processing power requirement is satisfied, but bandwidth usage increases due to continuous video streaming

Engineering Contradiction:
Improveaction recognition accuracyVSAvoidbandwidth usage
Core Design Contradiction:
Measurement precisionVSLoss of energy

Solution Approach 1:

The camera extracts and transmits only the essential information (action candidate detections and relevant video segments) to the server, rather than continuously streaming the entire video feed. This extraction approach dramatically reduces bandwidth consumption while providing the server with sufficient data to perform accurate action recognition verification.

Inventive Principle:
Principle #2Taking out (Extraction)

3Loss of energy

If a first action recognition algorithm is applied on the camera to detect action candidates, then processing load on the camera increases, but the algorithm can trigger selective video transfer to reduce bandwidth

Engineering Contradiction:
Improvebandwidth usageVSAvoidprocessing load on camera
Core Design Contradiction:
Loss of energyVSDevice complexity

Solution Approach 1:

The first action recognition algorithm performs preliminary detection of action candidates on the camera before any video transfer occurs. This preliminary action filters the video stream so that only segments containing potential actions are transmitted to the server, significantly reducing the total bandwidth required for action recognition processing.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3321844B1Action recognition in a video sequence
Publication Date: 2021.04.14 AXIS
  • EP3321844B1 patent drawingFigure 1
  • EP3321844B1 patent drawingFigure 2

AI summary

The present disclosure relates to a system for action recognition in a video sequence. The system comprises a camera (100) configured to capture the video sequence and a server (200) configured to perform action recognition. The camera comprises an object identifier (124) configured to identify an object of interest in an object image frame of the video sequence; an action candidate recognizer (126) configured to apply a first action recognition algorithm to the object image frame and thereby detecting presence of an action candidate; an video extractor (127) configured to produce action image frames of an action video sequence by extracting video data pertaining to a plurality of image frames from the video sequence, wherein one or more of the plurality of image frames from which the video data is extracted comprises the object of interest; and a network interface (130) configured to transfer the action video sequence to the server. The server comprises an action verifier (210) configured to apply a second action recognition algorithm to the action video sequence and thereby verify or reject that the action candidate is an action. Also a method for action recognition in a video sequence is presented.