A research team at the University of Zürich published a paper detailing how they got drones to fly on street-level in a safe manner.
Although companies like Amazon announced years ago that they will deliver packages using drones with their Prime Air system, few drones are flying between buildings. A big reason for this is that path-planning around obstacles on street-level is a difficult task for drones. To overcome this challenge, the researchers attempted to create a program that can estimate a sensible direction for the drone to travel in while simultaneous estimating if the drone is close to an obstacle. The idea is similar to that of drones navigating forest trails autonomously, an idea the same group published in 2016.
To predict the steering angles and possible collisions, the researchers created a deep neural network. The neural network architecture they train is a so-called residual network. It produces a steering angle to navigate the drone itself, and a collision probability so the drone can recognize dangerous situations and react to them.
An important contribution is the way they gathered the training-data. Gathering data using trained pilots in all these situations would consume a lot of time. Instead, the researchers trained their neural network using data from cars and bicycles. The dataset used to predict a good steering angle is the Udacity training set. Each image also contains the steering angle of the car which allows the drone to learn by behavioral cloning: try to predict the steering angle of the car using a neural network. They gathered their own collision dataset by mounting a camera on a bike and driving towards cars and other obstacles. This way their neural network is able to predict if the drone is close to an obstacle or not.
The researchers describe their results in a YouTube video, published their results in a paper, and published their code on GitHub.