Leveraging Edge Based Deep Neural Networks for Road Damage Detection

With the increasing number of casualties resulting from road damages, posing a serious threat to every individual and society; imperative action in solving the issue must be given primacy. Over the years, various approaches and solutions have been presented by researchers with a noticeable pattern-every breakthrough solution solves the shortcomings of its predecessor. With the recent advancements in image classification techniques through deep neural networks, it is now possible to classify road damages with high accuracy. However, one disadvantage of employing this technology is the considerable latency associated with running machine learning models in the cloud in real-time. Considering all these complexities, this paper presents an edge computing framework that runs efficient deep neural networks thereby reducing the latency inherent in the previous approaches. This solution can bring drastic changes to road maintenance by providing crucial information at opportune times thereby substantially reducing road accidents. Using transfer learning-based models, an F1 score of 0.64 was achieved for the RDD2020 dataset.