Project title

Ball Tracker, a ball detection and tracking pipeline using a pre-trained YOLO v3 object detector and OpenCV Tracking API

Motivation

I build this project for learning computer vision and also for fun. Building a detection and tracking pipeline from scratch is something I have always wanted to do since I started studying OpenCV. Also, I feel like there's something really profound in the act of drawing bounding boxes on images, sort of like a technology that comes out of a sci-fi or video games.

Tech/framework used

• Python
• OpenCV
• Darknet YOLO v3

Features

• Accurate object detection using YOLO v3.
• Fast and optimized tracking with OpenCV Tracking API
• Ease of customizing detection/tracking settings
• Easier to run

How to use?

First and foremost, download the yolov3.weights and put in in the models/ folder. Then simply run python3 main.py [flag] [path_to_video]

Demo

Limitations

• Inference takes too long, resulting in jagged video feed. This is because I'm using a detector that's built for 80 classes to detect only 1 class.
• Low frame rate if no object is detected.
• Lacking the ability to write and save the video that contains the bounding boxes.
• Lacking the ability to detect and track multiple objects.
• Only allow tracking objects that have > 90% detection confidence.

What's next?

• Speed up inference.
• Build my own neural net for object detection.
• Implement a Kalman filter to track the detected object and compare the performance with OpenCV Tracking API, cuz why not?
• Implement multi-object detection and tracking.
• Build a ball-tracking robot. Why? Because I'm a Mechatronics Engineering student, I want to integrate everything I have learned into a robot :D

Credits

Big thanks to @spmallick, @jrosebr1 for providing great learning resources. Huge credit to @pjreddie for his work on the YOLO v3 object detector. The YOLO v3 model used in this project can be found here