Overview / Usage

In April 2015, Nepal was hit by massive earthquake of around 7.3 Richter Scale. Which killed nearly 9,000 people and 22,000 were injured. It was one of the biggest earthquake to ever hit in Nepal since 1934. Mainly rural parts of Nepal were affected and lot of houses were destroyed. Because of this hundreds of thousands of Nepalese were made homeless with entire village flattened, across many districts of the country.

As an effort to compensate the building/house losses, Nepal government started a program to distribute the money of worth Rupees 300,000 to the victims of rural part of Nepal who has lost their houses. After careful survey of each of such victims, the compensation money were distributed among the victims. Many victims were able to re-construct their houses because of this program. But there was no way to monitor and re-assure if the houses were being made as per the standard set out by the government.

The only way to do such task was to deploy government employees and manually check every houses in rural part of Nepal. Which will be time and resource consuming task. This project aims to solve that. Using Satellite imagery data and a deep learning model, this project will identify, detect and label the houses which are not under the standard set out by the government, resulting in automatic monitoring and control of re-construction of such houses from all over Nepal.

Methodology / Approach

For Object Detection, Single Shot MultiBox Detector (SSD) model is used.
VGG 16 part of the model was loaded with weights obtained by training on ImageNet Dataset
SSD model then was trained on PASCAL VOC dataset which has images of common objects in them
Transfer learning was performed on small dataset of annotated satellite images of rural houses in an effort to maximize the accuracy of the model.

Technologies Used

Intel® optimized Tensorflow*
OpenVINO™ Toolkit
Intel® AI DevCloud
Intel® Movidius NCS
Intel® nGraph™