Object detection in UAV imagery is an important task with many downstream use-cases such as inventory management, and anomaly detection. The Vimana platform supports object detection both on the processed maps (ortho-mosaics), and the raw UAV images. In this article, we present two case-studies to describe these capabilities.
Detecting vehicles in construction sites is important both for inventory (track earth-movers, etc.) , and for safety checks (eg. cars or trucks parked too close to pits). We employ a multi-class object detection model (Mask-RCNN) that can identify and label vehicles into categories (eg. trucks, earth-movers). Further, for large vehicles such as earth-movers, the detection also provides a tight outline of the vehicle, allowing for accurate geo-referencing of the vehicles.
This feature is available on both ortho-mosaics, and raw images, and the output is overlayed on the images. In addition, one can also edit or annotate the individual detected items for further processing.
Solar Panel Detection
Detecting solar panel modules in power-plants helps both in estimating power output, and detecting anomalies such as incorrectly oriented or damaged modules. Our solution allows accurate detection (geo. referenced up to individual modules), while requiring minimal training examples (less than 10 annotations of each type of module). Additional post-processing helps identify both the total area covered, and the orientation (incline, and direction) of the individual modules allowing easy location of faulty modules.
Here, we use a segmentation model (U-net) along with custom logic to identify the shapes of the individual modules. This approach allows the model to be trained on images with only a few annotations of the modules, even if there are many modules visible in the images.
These features are available in the Professional, Team and Organisation plans of our platform. Please contact us for further information.