jalotra / queue-detection Goto Github PK

View Code? Open in Web Editor NEW

A naive Algorithm that uses People Detection and Convex Hull as subroutines to solve this problem: "Given an image of people standing randomly, how many people can be a part of some queue {Q}. A queue {Q} is basically a line(y = mx + c) that minimises a custom loss function. "

Shell 7.20% Python 92.80%

queue-detection's Introduction

Queue-Detection

Objective and Approach:

To find a possible queue vector which is basically a line or (y = mx + c) using People Detection as a subroutine and then answer this question : "How many people are standing in the queue".
Assumption Taken : "Whatever be the camera angle is the queue is always a planar surface in a 3D world".
Finds the bounding boxes around people using Yolo.
Of all those bounding boxes the code tries to find the best params(m, c) such that maximum people lie on or some delta(eps) along this line.
Takes 90% of all the points to remove outliers.
Finding params(m, c) is done by solving an optimisation problem that uses Convex-Hull as a subroutine.

For more details refer the Presentation that I made : Presentation

Current Tasks

Yolo from here : https://github.com/ultralytics/yolov5

To use the YOLO Algorithm to detect all the bounding boxes that are persons.
Reading the Source Code of Darknet, learning about Yolo Layers and the implementation in Pytorch.
Maybe Fine Tune the Model to only output bounding boxes around humans. Something like this : https://www.codeproject.com/Articles/5283660/AI-Queue-Length-Detection-Counting-the-Number-of-P

Setting Up the Development Arena

You must have Conda installed, since it provides containerisation we dont have to worry about building and running on Different Platforms/ OSes.
First run "bash setup_env.sh" -> This will create a Queue_Detection conda environment.
Run conda activate Queue_Detection to activate the newly made env.
Next run "bash setup.sh" -> This will install all the related dependencies.
In case of error, read the *sh files and figure it out yourself.