ElanGuard Readme

Demo Codes (GPU REQUIRED)

• gui_multicam.py
  Example of building up a muti-camera view GUI
  

• gui_singleview_personReID.py
  Example of tracking and identifying guest in single camera view

• gui_singleview_personReID_faceReID.py
  Example of tracking and identifying employee in single camera view

System Design

• Flow Chart

• Edge Overview

• Edge Dataflow

• Server Overview

• Server Dataflow

Install

1. Install Anaconda

Anaconda is highly recommanded Otherwise may have to install the dependencies in environment.yml manually

For Anaconda, simply create the environment with environment.yml : conda env create -f environment.yml
(For details, see Creating an environment from an environment.yml file.)

After installed, enter the environment with source activate ElanGuard

2. Install some of the modules if needed

• Yolov2
  cd ./src/detector/darkflow/
pip install .
  https://github.com/thtrieu/darkflow

• Yolov3
  pip3 install yolo34py-gpu
  https://github.com/madhawav/YOLO3-4-Py

• Convolutional Pose Machine
  cd ./src/detector/tf-pose-estimation/
pip3 install -r requirements.txt
  https://github.com/ildoonet/tf-pose-estimation

• ArcFace
  pip install mxnet-cu80 --pre
pip install easydict

3. (Optional) Install [Spyder IDE]

source activate ElanGuard
conda install spyder
spyder

(https://en.wikipedia.org/wiki/Spyder_(software)) for developing
Spyder provides an easy way to visualize the data structure

Download

models, videos can be downloaded in Google Drive
( Please contact us if permission for access is needed )

Computation Consumption

Module	Algorithm/Model	Code Version	Maintainer	Last update	Data Memory	GPU Memory	Computation Time(ms/frame)
Detector	MTCNN	1.0		2018.06.04	:---------:	:--------:
Detector	MTCNN_Batch	1.0		2018.06.04	:---------:	:--------:	50
Detector	MTCNN_Batch_Fixed	1.0		2018.06.05	:---------:	:--------:	10
Detector	Yolov2	1.1		2018.05.21	:---------:	:--------:	4
Detector	Yolov2+MTCNN	1.1		2018.05.21	:---------:	:--------:	30
Detector	Yolov3	1.0		2018.06.04	:---------:	:--------:	90
Detector	Yolov3+MTCNN	1.0		2018.06.04	:---------:	:--------:
Detector	MaskRCNN	1.1		2018.05.21	:---------:	:--------:	30
Detector	ConvPoseMachine	1.0		2018.05.09	:---------:	:--------:
Tracker	TripletTracker	1.1		2018.06.19	:---------:	:--------:
Embedder	Triplet_Embedder	1.0		2018.06.19	:---------:	:--------:	With Batch: 1.4/person
Embedder	ArcFace_Embedder	1.0		2018.06.19	:---------:	:--------:	With Batch: 3/person ; Image: 18/person
F_Identifier	ArcFaceReid	1.0		2018.06.19	:---------:	:--------:
P_Identifier	TripletReid	1.0		2018.06.19	:---------:	:--------:

Todo List

• Skip Verification for faces
• GPU resource allocation
• Batch based embedding haven't check max size
• Cross view tracking
• Filter out poor feature with mean and std
• OpenPose
• Inter-view threshold may be calculated by tracker
• MTCNN optimize (send batch )
• MTCNN detect wrong face when inference in Yolo
• Server database access speed

Modifications

6.19 Multi processor version

We are always welcome for questions and suggestions !!

References

Module	Source
Yolov2	https://github.com/thtrieu/darkflow
Yolov3	https://github.com/madhawav/YOLO3-4-Py
MTCNN	https://github.com/kpzhang93/MTCNN_face_detection_alignment
ConvPoseMachine	https://github.com/ildoonet/tf-pose-estimation
MaskRCNN	https://github.com/matterport/Mask_RCNN
IOUTracker	https://github.com/bochinski/iou-tracker
Triplet_Loss	https://github.com/VisualComputingInstitute/triplet-reid
Arc_Face	https://github.com/deepinsight/insightface
Align_Reid	https://github.com/huanghoujing/AlignedReID-Re-Production-Pytorch