Urban Safety Prediction Using Context and Object Information via Double-Column Convolutional Neural Network

Recently, various studies are performed on theanalysis of urban environment (e.g. safety, hygiene, economyand population). Particularly, the prediction of urban safetyhas received a lot of attention in society. In this paper, wepropose safety prediction of urban place using ConvolutionalNeural Network. To predict more accurate safety score, we usethe fusion of context and object information. To do this, weadopt Double-columns Convolutional Neural Network whichconsist of one column is context information extraction column, the other object information extraction column. Contextinformation is extracted from re-sized whole image and objectinformation is extracted from highest saliency score patchof object saliency map. We trained our prediction modelusing Place Pulse 1.0 dataset. To evaluate the performanceof our prediction model, we compared with SVR model andAlexnet model using RMSE. Also, we analyze correlationground truth safety scores and predicted safety scores. Fromour experimental results, our prediction model showed thebest performance (RMSE of 0.7403 and pearson/spearmancorrelation coefficient of 0.903/0.900).