Overview

This project is built from scratch with minimal use of black-box libraries.
The motivation for this project was child safety, when left alone at home, are in danger
of speedy moving objects and an alert can be sent to the parent. The algorithm can also
detect any object (of any size, shape, color, both living and non-living objects). It can also
be used to detect elderly people falling down. This application can particularly prove to be
useful in regions prone to earthquakes, and also at construction sites.


This project integrates a static background and utilizes a single point perspective camera view.
The algorithm ensures to only detect objects moving with a speed greater than a certain threshold.

Design

The algorithm's computational demands are significantly reduced as it avoids the use of neural networks.
Consequently, it can swiftly track objects in real-time, proving particularly advantageous when rapid
detections are essential in this specific application.This project is camera type independent where
the resolution doesn’t affect the algorithm and image processing and even low resolution
video can perform the task efficiently. Each step can be depicted visually, enhancing
transparency and facilitating better comprehension and understanding of each individual
action.

Algorithm

This algorithm can be broken down into four main steps, namely motion masking,
connected component labelling, object id and tracking, and speed filter. In motion
masking the pixel intensities are compared between frames, in connected component
labelling different objects get different labels, then a unique id is extracted for unique
objects and are tracking by calculating the similarity score of each object. Lastly, a speed
filter is applied to only track those objects which are not just moving but having a
speedy motion.

Demonstrations

Future Scope

Machine learning can be used for object similarity
Eliminate shadow tracking
Use other technique like edge detection like OpenCV Canny Edge Detection
Instead of bounding boxes, obtain object shapes by using segmentation techniques

Youtube Link

Video Presentation Link

Project Partner

Charles Cheng