Road dangers have caused numerous accidents, thus detecting them and warning users are critical to improving traffic safety. However, it is challenging to recognize road dangers from numerous normal data and warn road users due to cluttered real-world backgrounds, ever-changing road danger appearances, high intra-class differences, limited data for one party, and high privacy leakage risk of sensitive information.To address these challenges, in this thesis, three novel road danger detection and warning frameworks are proposed to improve the performance of real-time road danger prediction and notification in challenging real-world environments in four main aspects, i.e., accuracy, latency, communication efficiency, and privacy.Firstly, many existing road danger detection systems mainly process data on clouds. However, they cannot warn users timely about road dangers due to long distances. Meanwhile, supervised machine learning algorithms are usually used in these systems requiring large and precisely labeled datasets to perform well. I also would like to extend my sincere thanks to Prof. Dr. Xiaoming Fu, Prof. Dr. Marcus Baum, and Prof. Dr. Shengjin Wang for their precious time, guidance, and help during my Ph.D. study. I am very grateful to Prof. Dr. Thar Baker since he never hesitates to provide me feedback, guidance, and help.