项目启动，本项目主要综合行为识别、目标检测、物体识别、文本情感分析进行开发.由于我负责行为识别和目标检测这一模块，只整理自己的工作内容

Vision

1
2
Through image, video recognition, text recognition, the
students' words and deeds are processed by the system to determine whether it has adverse effects on public civilization.

Environment Support

  1. Keras 2.0/2.2
  2. Tensorflow 1.2
  3. Pytorch

    Main References

    I GET THE set of papers from HERE
    VIDEO TO TEXT
    IMAGE CAPTION
    GET THE NAME FORM DETAIL F-CNN
    CVPR

    Two-stream

    The development of my Action-Recgnization module is based on Two Stream 《Two-StreamConvolutional Networks for Action Recognition in Videos》.

Reasons of using Two-stream

The Action-Recgnization is developed on the way of Two-Stream recent years,And researchers have come out many papers on IEEE SCI and others.but the main reason is that I did Video caption last year, both of them are Analyzing Video Infomation, I want try other algorithm to finish my project in higher quality(get higher score).

Procedure

Graphviz

  • using dot generate this picture
    1
    2


the basic of Two-Stream is The Fusion of spatiotemporal information in a dual stream network.
or
The KEY POINT is the better Fusion of spatial and temporal features

  • The interaction between layers within a single network, such as ResNet/Inception.
  • between dual-stream networks, including the exploration of different fusion methods. It is worth considering the structure of ResNet and connecting the dual-stream network.

This project use the second method.

Spatial network

It mainly captures important object features in video frames.

Time series network

both of them : finetune the ImageNet

Document-Quality Attributes

Each contains:

1
2
3
4
5
6
7
8
9
10
.
└── Avaliability(Quality Attributes eg.)
     ├── Scenario
     ├── Stimulus Source
     ├── Stimulus
     ├── Artifact
     ├── Environment
     ├── Response
     ├── Response measure
     └── Tactics
Avaliability Performance Modifiability Usability Security Testability
Scenario Can not identify bad behavior The exported files are shown well New demands&Structural optimization Customers want to export statistics file easily and need a reliable data Databases is intruded Unit testing
Stimulus Source System dependencies Customers Developers and Customers Customers Attackers Developers
Stimulus Can’t solve information of the video Exporting operation Customers Runtime(?) Sql injection&entitlement Unit testing each module
Artifact whole system UI system UI DBMS Code
Environment Windows/Linux x86_64/32 in Runtime environment Web browser Runtime environment Web browser Firewall&Encryption Runtime environment
Response Send a feedback to backend if can’t analyze the video from surveillance cameras;Retry if can’t export the list of score Export statistics files in 10s Extend and modify functions when come out a new demand Provide a easy-operated UI and reliable information Resist intrusion Each module passed the Test Cases
Response measure within 5min;within 5s within 10s All modules is extensible and under the control of the evaluation indexs The satisfaction of user Database is protected Developers
Tactics Retry Self-test Increase Resource Efficiency Split Module aaa warm backup Limit Non-determinism

Project Details

  • This project explores prominent action recognition models with UCF-101 dataset
  • Perfomance of different models are compared and analysis of experiment results are provided

To be continued