OpenVino For People Counter

Table of Contents

• 1. Introduction
• 2. Architecture Block Diagram
• 3. Prerequisites
• 4. Build
• 5. Usage
  • 5.1. Run
    • 5.1.1. Detection, Count and duration
  • 5.2. Performance analysis
  • 5.3. Bonus
• 6. To Do
  • 6.1. README
  • 6.2. Development

1. Introduction

The people counter application will demonstrate how to create a smart video IoT solution using Intel® hardware and software tools. The app will detect people in a designated area, providing the number of people in the frame, the average duration of people in the frame, and total count.

People counting applications can be used in a retail store, supermarket, shopping malls, metro station, airport. For example, Once a person is detected, We can follow the object through a shopping pipeline in a retail setting or track and collate behavior i.e customer traffic patterns and time spent at merchandising location.

2. Architecture Block Diagram

3. Prerequisites

To run the application in this tutorial, the OpenVINO™ toolkit and its dependencies must already be installed and verified using the included demos. Installation instructions may be found at: https://software.intel.com/en-us/articles/OpenVINO-Install-Linux or https://github.com/udacity/nd131-openvino-fundamentals-project-starter/blob/master/linux-setup.md

The below steps are tested on Ubuntu 16.04:

# Install OpenVino
wget http://registrationcenter-download.intel.com/akdlm/irc_nas/16057/l_openvino_toolkit_p_2019.3.376.tgz
tar -xvf l_openvino_toolkit_p_2019.3.376.tgz
cd l_openvino_toolkit_p_2019.3.376
sed -i 's/decline/accept/g' silent.cfg
sudo ./install.sh -s silent.cfg

# Install Nodejs and its dependencies
curl -sSL https://deb.nodesource.com/gpgkey/nodesource.gpg.key | sudo apt-key add -
VERSION=node_6.x
DISTRO="$(lsb_release -s -c)"
echo "deb https://deb.nodesource.com/$VERSION $DISTRO main" | sudo tee /etc/apt/sources.list.d/nodesource.list
echo "deb-src https://deb.nodesource.com/$VERSION $DISTRO main" | sudo tee -a /etc/apt/sources.list.d/nodesource.list
sudo apt-get update
sudo apt-get install nodejs

# System dep
sudo apt update
sudo apt-get install python3-pip
pip3 install numpy
pip3 install paho-mqtt
sudo apt install libzmq3-dev libkrb5-dev
sudo apt install ffmpeg
sudo apt-get install cmake

# Create a virtual env
python3.5 -m venv openvino_p3
source openvino_p3/bin/activate
pip3 install dlib paho-mqtt fysom imutils

# Project dep
# To-do: Create a requirements.txt for project specfic
pip3 install

4. Build

1. Clone the repository at desired location:

https://github.com/nullbyte91/openvino-people_counter_IoT_app

2. Configure the build environment for the OpenCV toolkit by sourcing the "setupvars.sh" script.

source /opt/intel/openvino/bin/setupvars.sh -pyver 3.5

3. Change to the top git repository:

cd openvino-people_counter_IoT_app

4. Install web server dep:

# For MQTT/Mosca server:
cd webservice/server
npm install

# For Web server:
cd ../ui
npm install

5. Model conversion (IR using model optimizer)

#Install MO Dep
pip3 install -r /opt/intel/openvino/deployment_tools/model_optimizer/requirements_tf.txt

If you are using caffe object detection model use requirements_caffe.txt, requirements_onnx.txt for onnx.

Based on initial experimental work, I decided to use SSD based object detection. Let try ssd_mobilenet_v2_coco and ssdlite_mobilenet_v2_coco.

# Downlaod ssd model from Tensorflow model zoo
# ssd_mobilenet_v2_coco
wget http://download.tensorflow.org/models/object_detection/ssd_mobilenet_v2_coco_2018_03_29.tar.gz

tar -xvf ssd_mobilenet_v2_coco_2018_03_29.tar.gz

mkdir ssd_mo_model

python3 /opt/intel/openvino/deployment_tools/model_optimizer/mo_tf.py --input_model ssd_mobilenet_v2_coco_2018_03_29/frozen_inference_graph.pb --tensorflow_use_custom_operations_config  /opt/intel/openvino/deployment_tools/model_optimizer/extensions/front/tf/ssd_v2_support.json --tensorflow_object_detection_api_pipeline_config ssd_mobilenet_v2_coco_2018_03_29/pipeline.config --reverse_input_channels -o ssd_mo_model

We can use --data_type to get quantized weights {'FP16', 'FP32', 'half', 'float'}

The IOU is better with Intel pre-trained model person-detection-retail-0013

python3  /opt/intel/openvino/deployment_tools/open_model_zoo/tools/downloader/downloader.py --name person-detection-retail-0013 -o .

# ssdlite_mobilenet_v2_coco
http://download.tensorflow.org/models/object_detection/ssdlite_mobilenet_v2_coco_2018_05_09.tar.gz

tar -xvf ssdlite_mobilenet_v2_coco_2018_05_09.tar.gz

mkdir ssdlite_mo_model

python3 /opt/intel/openvino/deployment_tools/model_optimizer/mo_tf.py --input_model ssdlite_mobilenet_v2_coco_2018_05_09/frozen_inference_graph.pb --tensorflow_use_custom_operations_config  /opt/intel/openvino/deployment_tools/model_optimizer/extensions/front/tf/ssd_v2_support.json --tensorflow_object_detection_api_pipeline_config ssd_mobilenet_v2_coco_2018_03_29/pipeline.config --data_type FP16 -o ssd_mo_model/FP16

Precision benchmark:

Table 1. Table Title

Model	FP32	FP16
ssd_mobilenet_v2	65M	33M
ssdlite_mobilenet_v2	18M	8.6M
person-detection-retail-0013	2.8M	1.4M

5. Usage

5.1. Run

5.1.1. Detection, Count and duration

1. Output on OpenCV window:

python3 main.py -m ssd_mo_model/FP32/frozen_inference_graph.xml -i resources/Pedestrian_Detect_2_1_1.mp4 -l /opt/intel/openvino/deployment_tools/inference_engine/lib/intel64/libcpu_extension_sse4.so

2. Output on Web Server:

Terminal 1: Start the Mosca server

cd webservice/server/node-server
node ./server.js

# You should see the following message, if successful:
Mosca server started.

Terminal 2: Start webserver GUI

cd webservice/ui
npm run dev

# You should see the following message in the terminal.
webpack: Compiled successfully

Terminal 3: FFmpeg Server

sudo ffserver -f ./ffmpeg/server.conf

Terminal 4: Run a application

source /opt/intel/openvino/bin/setupvars.sh -pyver 3.5
python3 main.py -m ../openvino-people_counter_IoT_app/ssdlite_mo_model/FP16/frozen_inference_graph.xml -i resources/Pedestrian_Detect_2_1_1.mp4 -l /opt/intel/openvino/deployment_tools/inference_engine/lib/intel64/libcpu_extension_sse4.so -d CPU -pt 0.6 -o WEB | ffmpeg -v warning -f rawvideo -pixel_format bgr24 -video_size 768x432 -framerate 24 -i - http://0.0.0.0:3004/fac.ffm

5.2. Performance analysis

Detection, Count and duration

Hardware configuration: i7-6820HQ CPU

Model	FP32 - Size	FP32 - FPS	FP16 - Size	FP16 - FPS
ssd_mobilenet_v2_coco_2018_03_29	65M	7.61	33M	8.01
ssdlite_mobilenet_v2	18M	11.02	8.6M	13.06

5.3. Bonus

Detection with high FPS using async with multithreading

cd detection

python3 main.py -i ../resources/Pedestrian_Detect_2_1_1.mp4 -m ../openvino-people_counter_IoT_app/ssdlite_mo_model/FP16/frozen_inference_graph.xml -l /opt/intel/openvino/deployment_tools/inference_engine/lib/intel64/libcpu_extension_sse4.so -pt 0.5

This implementation gives 90 FPS on the same hardware.

6. To Do

6.1. README

• ✓ docker compose for deployment

• ✓ Performance analysis

6.2. Development

• ✓ Try with different models

• ✓ Kalman filter based tracking

• ✓ dlib based tracking

• ✓ Fix labels issue

• ✓ supermarket use cases

• ✓ async with multithread implementation