ayan-cs / scene-reconstruction-edge-metadata

Reconstruction or Approximation of environment from edge data. Using pre-trained models, we are analyzing the incoming edge data (CCTV camera footage) and extracting the detected objects and person's pose. These metadata are sent to remote server, where using these metadata, the video scene is reconstructed or approximated on a Real-time manner.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Scene Reconstruction for Edge using Metadata

IMPORTANT : Due to organization policy, source code cannot be shared. The basic idea can be obtained from the respective README.md files.

Objective

Reconstruction or Approximation of environment from edge data. Using pre-trained models, we are analyzing the incoming edge data (CCTV camera footage) and extracting the detected objects and person's pose. These metadata are sent to remote server, where using these metadata, the video scene is reconstructed or approximated on a Real-time manner.

Installation of Necessary Libraries/Packages

There are two different implementations of the system using two different communication protocols. They are as follows,

  • Message Queuing Telemetry Transport (MQTT)
  • Asymmetric Message Queuing Protocol (AMQP)

Find the respective source codes in respective branches. The necessary libraries and/or packages are mentioned in the respective requirements.txt files. Use the command below to install necessary dependencies.

pip install -r requirements.txt

Additional Requirements for MQTT

You need to install the MQTT broker from Mosquitto or if you have Docker installed on your machine, then you can easily run the Dockerized Mosquitto which will act as your broker. The code does not required to be changed for any of these.

Regarding AMQP Implementation

For running the system, RabbitMQ has been used. RabbitMQ is the AMQP implementation on AWS platform. Create a RabbitMQ broker on AWS and note down the Queue id and other necessary attributes which is/are to be defined in the code.

Running the System

  • Clone the Repository.
  • Switch to the necessary branch with respect to the protocol.
  • Start running the broker.
  • Open two different Command Prompts.
  • On one prompt, run the Publisher.
  • On the other prompt, run the Subscriber.

Constraints and Future Improvements

  • The system is only limited to Single person pose estimation
  • Object detection using YOLO-v3 is quite slow and sometimes inconsistent
  • The Image superposition only works if the object is fully inside the frame (Camera's field of view)

About

Reconstruction or Approximation of environment from edge data. Using pre-trained models, we are analyzing the incoming edge data (CCTV camera footage) and extracting the detected objects and person's pose. These metadata are sent to remote server, where using these metadata, the video scene is reconstructed or approximated on a Real-time manner.