Drone-Assisted Inspection for Automated Accident Damage Estimation: A Deep Learning Approach

Abstract

Drones have been used in many application domains nowadays including traffic congestion control, weather information collection, disaster and rescue interventions, and surveillance operations. The drone adoption lies on their capabilities to collect images, videos as well as other sensory data from the air, stream this data to the cloud for processing, and analytics in order to derive important real-time decisions. In this paper, we propose a drone assisted inspection for accident damage estimation based on deep learning approach. Drones are automatically scheduled to visit the accident locations, and data is retrieved for further processing and analytics. We developed a two-phases damage estimation approach, where in the first phase we use deep learning approach to identify and classify objects from accident's images, and in the second phase we measure the size of damaged objects and we estimate the overall cost of the accident's damages. We evaluated our two-phase approach using data of various accidents, and the classification accuracy we have obtained vary between 0.79 and 0.94 and the accident's damage cost estimation most of time is 100% accepted by the expert.