GTC-DC 2019: Developing Altitude-Agnostic Computer Vision Algorithms for Edge Computation

GTC-DC 2019: Developing Altitude-Agnostic Computer Vision Algorithms for Edge Computation

Dean Teffer, Slingshot Aerospace; Brian Williams, Slingshot Aerospace
We’ll discuss a method of developing altitude-agnostic computer vision algorithms. Performing object detection and classification on a sensor presents more challenges than cloud-based machine learning. We trained a semantic segmentation convolutional neural network (CNN) to detect buildings from high-altitude aircraft and satellite imagery. We adapted the CNN to perform in-flight processing on a Jetson TX2 onboard an unmanned aerial vehicle (UAV) using TensorRT optimizations. Despite the differences present in low-altitude footage, we’re able to perform well against UAV footage with limited fine-tuning using a sequence of normalization and preprocessing stops. The resultant algorithm can robustly detect buildings across a wide span of altitudes.