MwansaMwango / final-project-computer-vision

BizzWatch - An app that analyses camera/video footage to audit business revenue and customers behaviour.

BizzWatch

An app that analyses camera/video footage to audit business revenue and customers behaviour.

Deployment

Demo Live_link

Motivation

CCTV has been the goto solution to prevent employee or internal theft. Business managers can't be present at all times and are too busy to review hours of video footage. Employee theft continues to hinder profit margins for business owners, especially in most eveloping countries with average annual income is under US$ 3,000!

Solution

BizzWatch is a solution that leverages computer vision and data analytics to detect, audit and predict revenue/customer demand.

User Story

BizzWatch allows a user to upload a video with business rules/logic with options such as number of frames to skip and target class (e.g. person, car or bicycle) for object detection. The video is then split into frames / images that are fed into the pretrained model to perform detection. Thereafter, the detect button runs the inference / object detection. The results are displayed in a carousel of images showing inference bounding boxes and total count of objects. Additionally, sales/revenue vs. people count bar chart visualisation is generated according to the business rules/logic csv file data. Note: File Upload and online inference has been disabled due to Heroku free-tier limitations. Also dummy data is used for MVP sales / revenue to demonstrate how people counting can be correlated to sales/revenue over time.

High Level Design

Technology Stack

Front End

• HTML
• CSS
• Javascript
• Bootstrap

Back End

• Python
• Flask
• REST/API

Machine Learning

• OpenCV
• Pytorch
• Retinanet_r101_fpn
• ResNet 101 backbone Neural Network
• Pretrained COCO Weights

Benefits

• Fraud Detection - object detection vs. ground-truth / prediction
• Sales / Customer Predicion - using influencing factors/features
• Resource Optmisation - staff, goods and services

Future Works

• Use Streamlit for Front End
• Cloud deployment - paid-tier, to allow upload of video and cloud inference. Current deployment on free-tier Heroku has limited compute and storage.
• Human Pose Estimation to identify activity
• Dwell Time - time object spends in a given user defined zone
• Zone Region of Interest - specify exact areas of detection
• Real Time Video Streaming Inference - RTSP enabled IP camera

Development Time

• This project / minimum viable product was developed in an intense 2 Weeks before deadline of final bootcamp project.

Author

Mwansa Mwango

Acknowledgement

• Simple MMDetector
• Case Study - Crowdedness at the Campus Gym (Random Forest Regression)
  • Crowdedness at the Campus Gym Dataset