Sage Pedestrian Lite

Description of repo content

---

Students: Francisco Lozano, Justin Derus, Wesley Kwiecinski
Mentors: Pratool Bharti, Michael Papka

Useful Sources:

• Ported from https://github.com/ddiLab/SagePedestrian
• Open CV Image Writing: https://docs.opencv.org/3.4/d4/da8/group__imgcodecs.html (Python is below c++)
• Open CV Line Drawing Documentation: https://docs.opencv.org/4.x/dc/da5/tutorial_py_drawing_functions.html
• Object Detection Tutorial: https://tensorflow-object-detection-api-tutorial.readthedocs.io/en/latest/auto_examples/plot_object_detection_saved_model.html
• Deep Re-id: https://github.com/KaiyangZhou/deep-person-reid

---

STEPS TO GETTING THE CODE UP AND RUNNING (on SageEDU nodes):

Prerequisites

• Acquire a SageEDU node and bootstrap it following the SageEdu repo. This will set up the operating system, python, virtual environments, github, and the camera.

• Acquire jetson-inference to do object detection following this video

Creating the virtual environment

1. Following the documentation on SageEdu create a pip environment by running this command virtualenv ./PedLite

2. Activate the environment by running source PedLite/bin/activate

Note: Stay in this virtual environment in the preceding steps. Whenever you want to detect pedestrians you have to activate this environment

3. Clone this repo to your nano git clone https://github.com/ddiLab/SagePedestrianLite.git

Downloading required pip libraries & apt packages

Download the required pip libraries by running pip install [library name]

Libraries to install

• numpy
• cv2
• shapely (you must download libgeos-dev first)
• sympy
• recordtype
• pascal_voc_writer

Note: When downloading pip libraries be careful with using sudo. Using this will cause the package to NOT be downloaded in your virtual environment, but rather in the global python directory

Download the required apt packages by running sudo apt-get install [package name]

Packages to install

• libgeos-dev

Torch Libraries

Lastly, you need to install the torch libraries, but in the jetson nano you must do this using pre-built PyTorch pip wheels. To do so, follow this forum. Make sure you download the correct combination of Pytorch and torchvision based on your python version. Pay attention to what version of Pytorch gets download some of those pip wheels will result to another version getting downloaded. Finally, if you mismatch the wrong versions of pytorch and torchvision you will get an error. To install the right versions make sure you uninstall the previous version you downloaded since the install script will not do so.

Download Deep-Person-reid to do Re-identification

1. cd into this repo in your computer and clone this repo

   git clone https://github.com/KaiyangZhou/deep-person-reid.git
   

2. Make sure you have your virtual environment activated source PedLite/bin/activate

3. cd into the deep-person-reid directory that was just created inside the SagePedestrianLite directory

4. install dependencies by running this command pip install -r requirements.txt

5. install torchreid python setup.py develop

Note: Module 'disutils' has no attribute 'version' is a know bug in pytorch so you have to run this command to fix it pip install setuptools==59.5.0

Reference their Github repo for more information Deep-Person-Reid

Modifying the Armcrest Camera stream

To modify the camera stream go to your internet browser in your Jetson nano and type in the camera's ip address in the search bar reference SageEdu on how to find your camera's ip address. After that is done set up a new password after logging in with the default log in creds:

Username: admin
Password: admin

Once you are logged in you can modify the camera stream such as the timestap.

setup > Camera > video... then select overlay to modify the timestamp

Setting up your Sql Lite Database

After the images are analyzed the python script saves the pedestrian's coordinates in a database sitting in the jetson nano. To download sql lite in your EDU node run this command sudo apt install sqlite3. Go into the sqlite3 command line and create a database called pedestrian_detections and the tables specified in SagePedestrian.sql. Once that is done, put the .db file in this directory (i.e. SagePedestrianLite/)

Runnning the Python Scripts

TakePictures.py

Before you run this script you are going to have to change the username, password, and camIP variable to match your camera details. This script will take pictures and organize them by date in the image folder until you press ctrl + c. Before you run any other python scripts, this script must be ranned first. Essentially, the following scripts analyze the pictures this script took.

Setting up a new environment

The scripts currently uses /Images/env/Env1.jpg as a generic image for overlaying the pedestrain's trajectory. To set up the script for a new environment take a picture generated from TakePictures.py and move it into the /Images/env/ folder. After this is done change the Env_Image global variable in plot_lines.py to the name of the new .jpg file sitting in /Images/env/.

The rest of the python scripts

1. Make sure that the .jpg files for the day you run exist in this directory Images
2. To run the pedestrian_detection script STANDALONE, follow this format: "python pedestrian_detection.py [hour_min] [hour_max] [date1, date2, ...]"
   • Where hours are optional, dateN = yyyy/mm/dd
   • Otherwise running plot_object detection will call pedestrian_detection and plot lines.
3. If not running pedestrian detection alone, use python plot_object_detection.py, which will use the current date.
4. The output of the pedestrian_detection.py script will be found in /image_label_xmls/crosswalk_detections/(ran_date)

TROUBLESHOOTING

This material is based upon work supported by the National Science Foundation under Grant No. OAC 1935984.