Counting cars in Dubai from Satellite Images
An object detection approach to object counting on Satellite Images.
We train an object detection model to detect cars on google maps satellite images. We use this model to count the number of cars in Dubai.
In this article, we discuss our experience building an deep learning based object detection model that can detect cars on satellite images. We created a large dataset of google map’s satellite image dataset for Dubai and used the model to detect and count the number of cars in Dubai.
Results:
Total Number of cars counted in Dubai: ~298,000
- Here all 4 wheelers are labelled as “cars”
- The images were dowloaded on September 2021
- Needless to say, this approach is prone to errors — model errors, vehicles in sheltered parkings not detected etc.
Approach
- Download google maps tiles :
Hacked this awesome github repo to download google maps data as tiles: https://github.com/AliFlux/MapTilesDownloader
Downloaded 300k+ tiles of dimensions 256 x 256 pixels at zoom level 19
2. Sampled regions of the map for manual labeling. :
Try to make the regions as diverse, and geographically different as possible.
3. Stitch tiles and resample to 800 x 800 pixel images + manual label:
Use LabelImg for manual labeling
4. Train object detection model:
Training with Faster RCNN — Detectron 2 implementation
5. Infer on full set of tiles of Dubai
6. Count number of detections
Shoulders of Giants
- Detectron2: https://github.com/facebookresearch/detectron2
- MapTilesDownloader: https://github.com/AliFlux/MapTilesDownloader
- LabelImg: https://github.com/tzutalin/labelImg
- Keppler gl: https://kepler.gl/
- QGIS: https://qgis.org/en/site/
About The Authors
Noufal — Data Scientist, Dubai Holding -UAE. https://www.linkedin.com/in/kvsnoufal/
M. Sirajul Huda — Data Scientist, Bosch Engineering — India. https://www.linkedin.com/in/muhammed-sirajul-huda-k-6687a457/