ObjectDetection-v1

Backup of Old Project (June 2023)

This is a backup of an old project that focused on object detection and tracking over videos using YOLOv8. The project was based on the following tutorials:

• Murtaza's Workshop - Robotics and AI's Object Detection Tutorial
• Computer Vision Zone's Object Detection Course

The project was written in Python and uses the YOLOv8 object detection model.

Execution

Start a new virtual environment:

python3 -m venv venv
source venv/bin/activate # linux or windows WSL
.\venv\Scripts\activate # windows cmd

Then install the requirements:

pip install -r requirements.txt

Finally, run the jupyter notebook Object Detection.ipynb within the virtual environment.

About the Project / Key Features

Static Image Detection

• Applying YOLOv8 on static image given by path specified in the "Parameters" section.

Video Detection

• Applying YOLOv8 on video given by path specified in the "Parameters" section.

Instance Tracking with Abrewlay Sort

• Instance identification with Abrewlay Sort library for tracking objects.

Notes

• The color of the box indicates the INSTANCE ID

Problems

1. Inconsistent IDs: Occurs under occlusion or label changes
   • Observe that the id of the truck 457 on the rightmost lane (cyan, with label "car") changed to id 473 (purple, with label "truck")
   • Indicates that the library cannot provide a consistent ID for the same object across frames
2. Multiple bounding boxes for the same object
   • As artifacts of the original YOLOv8 detection (due to using the nano model)

Custom Tracking Algorithm

• Instance identification with custom-implemented algorithm for tracking objects.

Goals

• Implementing a version of the object tracking algorithm that is more resistent to lost frames and flickering compared to the Abrewlay Sort library.

Features

1. Label Accumulation
   • Can accumulate labels from previous frames.
   • The ids (colors of the frames) of the objects are more consistent, which can be checked visually
   • The same truck now got a consistent id (no change in color indicates that)
2. Bounding Box Merging
   • Finds the closest bounding box to the previous frames (stored in a dictionary)
   • Dictionary is cleared after a certain number of frames
   • Only bounding boxes within a certain change in size and aspect ratio are merged by overwritting the same instance id

Car Counter

• Simple algorithm for counting cars across the line by masking the image before detection.
• Utilized the algorithm in sectino 3.5.2 for tracking objects to ensure that a car is not detected twice.