I. Introduction
In autonomous driving tasks, based on the road condition images acquired in real time, pedestrians or vehicles need to be identified and located to support the intelligent autonomous transport system for its decision making. Existing methods are able to accomplish the requirements of detecting road conditions under most circumstances (e.g., sufficient light, high distinguishability), but in adverse environments (e.g., night, rain and fog), pedestrians and vehicles with similar characteristics to the environment are less distinguishable and can easily be overlooked by detection system, thus posing a serious threat to traffic safety.