Inspection of City Infrastructure Via Peripheral Perception

Overview 

One of the main tasks of any city government is to keep infrastructure such as roads, signs, accessibility ramps, and sewers in perpetual working order. The first step of maintenance is the timely identification of a problem. However, many infrastructure defects are left undetected and unattended for long periods of time.  

The City of Austin possesses several fleets of vehicles, which work regularly around the city to inspect infrastructure. We will investigate how to leverage those vehicles with state-of-the-art computer vision, robotics, and data science techniques to automate infrastructure inspection in such a way that is publicly acceptable, will reduce costs, and will increase the effectiveness of city maintenance efforts. 

Methods and Findings 

We have formulated the peripheral perception for infrastructure inspection problem as a novel Team Orienteering Coverage Planning with Uncertain Reward (TOCPUR) problem. As an abstraction for the specific infrastructure inspection via peripheral perception problem, the TOCPUR problem is also general and of interest to the planning and optimization community. TOCPUR relaxes existing orienteering problem and coverage planning problem’s assumption on known reward a priori, by adding uncertainty on the reward. The reward uncertainty is a better model for the infrastructure inspection task, because, for example, the location of the potholes on a city road network is usually unknown. This uncertainty further creates the exploration and exploitation trade off.  

Based on Mixed Integer Programing, we have developed a novel solution to the TOCPUR problem. We have generated a graph-based dataset, which can represent many different city road networks and made our solver public for the community’s use. We have deployed our method on a heterogeneous robot team of three mobile robots on UT Austin campus and shown that our method can achieve better coverage compared to a greedy baseline algorithm.