Submission Received "Special Recognition Award" for the phase 2 of the hackathon.
Members: Sundarakrishnan N, Sohan Varier, Tarun Bhupathi, Manaswini SK of RV College of Engineering
For more details on the implementation,check the report: Link
More details about the event can be found here: Link
A Dockerfile has been created to install necessary libraries including CUDA for the model to be able to use GPUs. The docker file can be built and run in a simple way.
Build: docker build -t username/imagename:version
Push: docker push username/imagename:version
Run: docker run --rm --runtime=nvidia --gpus all -v 'YOUR STORAGE MOUNT':/app/data username/imagename:version python3 app.py input.json <output-folder-directory>
The run command takes the input and output json file to read,process ands save the results in. Build creates the docker container and push command is used to push it to the docker repository so that anyone can pull the image and run the same.
This folder contains all the files needed to run the pipeline
-
app.py : The main driver code that has to be run. Takes an input JSON file and output file directory as CLI arguments that provide the video paths and the path to the final output folder, where the images and matrices will be stored after processing.
-
best.pt : The most important file, our trained YOLOv8 model to detect the 7 classes of vehicles.
-
deep_sort_realtime folder : The DeepSort-realtime library that we downloaded locally, as we had to make minor changes to the source code to suit our needs.
- in
tracker.py, modified the code to return the classes of the detections as well, along with the unique ID.
- in
Only app.py is supposed to be run, the other files cannot run on their own.
Command to run: python3 app.py input.json <output-folder-directory>
<output-folder-directory> refers to the folder under which the final matrices and images are stored.
Format for input.json:
{
"Cam1": "/app/data/Cam_1_vid.mp4",
"Cam2": "/app/data/Cam_2_vid.mp4"
}
These are some other scripts used to ease the process of trainng and development but is not needed to run the framework.
-
extract_images.py : We used this script to extract images from the video downloaded from the dataset every n frames which we can set based on the number of images required.
-
auto_annotate.py : After making a basic model, we ran the extracted images through the model to annotate the images for us, and we would verify/edit the annotations. This script automated the annotation process and saved us a lot of time.
-
data_split.py : A simple script to split the images dataset into training, testing and validation sets.
-
stream.py : This code lets us view the YOLO model in action on a live video. It shows us the predictions being made in real-time in the video.
-
data.yaml : This file is used to specify dataset location during training, and holds the list of classes.
-
opencv_python_headless 4.10.0.84 : Used to read the video files and display them if needed.
-
ultralytics 8.2.58 : The library that contains YOLOv8, the model we used for vehicle detection and counting.
-
pandas 1.5.3 : The basic data structure used throughout the project, pandas dataframes.
-
Pillow 10.4.0 : Python's Image Library (PIL) that is used to view and manipulate images.
-
torchreid 0.2.5 : A re-id library used by DeepSort. This is the library used in our re-identification model.
YOLOv8 by Ultralyitcs is an open source, real-time object detection and image segmentation model.
labelimg is an annotation tool that provides features to draw and edit bounding boxes in the format required by YOLOv8.
DeepSort is a library that contains pre-trained re-identification models that we can make use of. It makes use of the detections done by the YOLO model, and checks whether any of those are re-appearing later.
- CPU - Core i5
- GPU - NVIDIA GTX 1650
- RAM - 8 GB
- SATA - 10 GB
- Around 1GB of GPU memory would be used for realtime inference.