A Novel Approach for Predicting and Understanding Road Danger in the Developing World: Deep Video-Classification of Roads in Nairobi, Kenya

With a road traffic death rate of 27.8 per 100,000 inhabitants in 2016, Kenya has nearly twice as much road fatalities as the world average. Hence, understanding the factors determining road danger is key and we are the first ones to attempt at achieving this goal by constructing a deep learning model entirely based on videos of several road segments from Nairobi. The best-performing model is a pre-trained Res-Net3D with shorter video clips which results in a test set accuracy of 50%. Comparing the results with the rest of the proposed models, we are able to infer that road danger is not only a function of quality of the road, but also of the density of road and pedestrian activity within a given timeframe. While similar models for video classification of daily activities reach an accuracy of 70%, we believe that given the increased complexity of our classification task (road danger), we fare rather well as a first pass. Click here to access the GitHub repo of this project. 1. Motivation and Related Work Providing traffic safety and lowering the rate of road accidents in Nairobi, Kenya is a major concern. With a road traffic death rate of 27.8 per 100,000 inhabitants in 2016, Kenya has nearly twice as much road fatalities as the world average (WHO, ongoing). Collaborating with the World Bank, we are the first ones to construct a deep learning model entirely based on videos of several road segments from Nairobi. The model allows us to analyze different road conditions and predict danger level of roads.There have been previous studies which used deep learning to evaluate the risk of traffic accidents such as Hébert, Antoine, et al. (2019), Chen et al. (2016) and Yuan et al. (2018). These studies have mainly focused on training models based on features such as weather, human mobility, road conditions and satellite images. However, we aim to develop models which process raw video data capturing traffic patterns in order to classify road danger level. Karpathy et. al. (2014) write the seminal paper on video classification using 3D CNNbased models. Later papers such as Abu-el-Haija et. al. (2016) and Diba et. al. (2017) come up with deeper and more advanced architectures and incorporate transfer learning in order to improve performance. The current state of the art model is given by Carreira & Zisserman (2017) who reach 80.9% accuracy on HMDB-51 dataset and 98.0 % on UCF101 dataset. Most of the papers use daily activity datasets, which is arguably a much easier classification task that the road danger evaluation that we aim to conduct.