Automated Attendance system using Deep Learning prinicples and Movidus Neural Compute Stick
- 0 Collaborators
Attendance taking has become an integral part in today’s collegiate world. It is necessary to inculcate a sense of discipline and responsibility in young minds to motivate them to succeed. Currently a significant amount of time is spent by faculties to take record of students present in the class. There is a time-accuracy tradeoff while recording attendance manually, namely if the faculty tries to speed up the process then there is possibility of false positives (read “proxy attendance”) and if the faculty tries to maximize accuracy, then he/she will stand to lose a lot of time in each class just taking attendance. This situation is further exacerbated in places such as India, where the average class sizes frequently exceed over 60 students per class. We propose a Deep Learning based solution to automatically mark attendance of students using face recognition techniques. ...learn more
Project status: Under Development
Intel Technologies
MKL,
Intel Opt ML/DL Framework,
Movidius NCS
Overview / Usage
The project aims to solve the Mass Surveillance Problem, and more specifically optimizes a specific subset of this problem by aiding teachers keep a track of all students currently attending a lecture. This research work will be useful a proof of concept for in-situ low cost inferencing in real time, and through the usage of Movidus Neural Compute Stick will shift the inferencing to the edge computing world, in an effort to deploy such technologies in remote areas and network congested areas.
Methodology / Approach
The project has multiple stages, the first being a training a two level convolutional neural network based architecture to identify the faces in a particular classroom. The first level would include a face detection algorithm to detect and crop all the faces in a given image, and the second layer would be the face recognition/ classification algorithm to identify and classify exactly to which person a face belongs to. We pipeline the decoupled architecture to enable real time inferencing of the faces.
The second stage is to demonstrate whether such a method could be placed inside a Movidus NCS, and to demonstrate if it would be possible to inference the images in real time
Technologies Used
Device: Movidus NCS, Ubuntu 16.04 Computer with optimized Intel MKL
Deep Learning Frameworks : Tensorflow, Keras
Image Manipulation: OpenCV