People Counter at the Edge
- 0 Collaborators
This application is designed to count the number of people and their duration in an environment. Its use cuts across different applications such as counting the number of people in a mall or using it to limit access of people to a place. ...learn more
Project status: Published/In Market
Internet of Things, Artificial Intelligence
Intel Technologies
OpenVINO
Overview / Usage
A lot of applications require the detection and counting of people in a space. One common method that is used to detecting people is through the use of motion detectors. However, motion detectors can be falsely triggered by moving objects of animals.
Another method is through the use of cameras. Using cameras reduces the possibility of false positives from moving objects/animals, as experienced by motion detectors.
Inferencing will need to be done on the images captured by the cameras. One way is to do the inferencing on the cloud - with associated latency, computational cost and privacy concerns. Another option is to perform the inferencing locally. This however requires sufficient computational resources on the part of the hardware.
This project utilizes the Openvino library from Intel to optimize the performance of a tensor object detection model (Ssd Mobilenet V2 Coco), to provide faster results on the edge, for detecting and counting people.
The project can be run be feeding the input from a camera to the people detection app running on an intel processor.
Methodology / Approach
This application makes use of Intel's OpenVINO toolkit. The model optimizer from the OpenVINO toolkit is used to convert the known layers of the tensorflow object detection model (Ssd Mobilenet V2 Coco) into corresponding internal representation, optimizing the model and producing a set of intermediate representation files.
The converted model is then integrated into an application that collects feed either from a camera or video input. The frames from the camera or video input is then fed into the model, which then provides the result.
The result is then preprocessed to count the number of people the model detected in the frame, in addition to the amount of time the people spent in the frame.
The pre-processed result is then sent over MQTT to a dashboard and can also be visualized through the use of FFServer.
Technologies Used
Python
Node.js
FFServer
OpenVINO
Intel Core i5 6500 TE