DeepStream Object Detection and Tracking

This repository is a simple project to demonstrate how to implement a pipeline for object detection and tracking using NVidia DeepStream. It also contains a simple implementation of a custom DeepStream plugin using python.

Requirements:

• DeepStream SDK = 6.0
• Python >= 3.6
• Python-Gst >= 1.14.0
• pyDS

1. System Design

This system is designed to do object detection and tracking on video stream/files and send results to a web API using post request.

System design illustrated below:

2. Installation

Before starting to install I highly recommend using Docker

Docker

To build

docker build --gpus all -t deepstream_tracking_sample:0.1 .

To run

docker run --gpus all --network host -i deepstream_tracking_sample:0.1

Ubuntu

There are three steps to install dependencies

1. DeepStream SDK

First you need to install DeepStream SDK, I assume you've already installed DeepStream SDK

The following dependencies need to be met:

2. Base dependencies

Ubunutu - 18.04 [using python-3.6]:

apt install -y git python-dev python3 python3-pip python3.6-dev python3.8-dev cmake g++ build-essential\
    libglib2.0-dev libglib2.0-dev-bin python-gi-dev libtool m4 autoconf automake

Ubuntu - 20.04 [using python-3.8]:

apt install python3-gi python3-dev python3-gst-1.0 python-gi-dev git python-dev \
    python3 python3-pip python3.8-dev cmake g++ build-essential libglib2.0-dev \
    libglib2.0-dev-bin python-gi-dev libtool m4 autoconf automake

3. Build python bindings

Gst-python:

git clone https://github.com/GStreamer/gst-python.git
cd gst-python
./autogen.sh
make
make install

After building and installing gst-python you need to set GST_PLUGIN_PATH variable to point to plugins, for this project plugins are located at the plugins folder

export GST_PLUGIN_PATH=$GST_PLUGIN_PATH:<project-path>/plugins

Pybind11:

git clone https://github.com/pybind/pybind11.git
cd pybind11
python3 setup.py install

PyDS:

git clone https://github.com/NVIDIA-AI-IOT/deepstream_python_apps/
cd deepstream_python_apps/bindings/
mkdir build && cd build

Ubuntu - 18.04 [using python-3.6]:

cmake ..
make
pip3 install pyds-1.1.0-py3-none-linux_x86_64.whl

Ubuntu - 20.04 [using python-3.8]:

cmake .. -DPYTHON_MAJOR_VERSION=3 -DPYTHON_MINOR_VERSION=8
make
pip3 install pyds-1.1.0-py3-none-linux_x86_64.whl

Python requirements:

pip3 install -r requirements.txt

3. Configuration

For the configuration there are some config files that are located in configs folder

3.1 Main Config

In this config file you need to provide project's main configs such as video stream uris, object detector's config, object tracker's config.
Example: main_config.yml

3.2 Object detector

Config file for nvinfer plugin
Example: pgie_config.txt

3.3 Object Tracker

Config file for nvtracker plugin
Example: tracker_config.txt

4. Run

To run the project after installing all the requirements just simply run:

python3 pipeline configs/main_config.yml

After running the command above, the pipeline starts to build and play, by detecting any object sendersink would send a post request to API, for testing results there is a simple Flask Rest-API that receives post request from sendersink.
To run the test API run the following commands:

pip3 install flask, flask-cors
python3 test/receiver_app.py

References:

• https://github.com/NVIDIA-AI-IOT/deepstream_python_apps
• https://github.com/GStreamer/gst-python