The images taken under varying lighting or adverse weather conditions exhibit different distributions in high-dimensional space, and make object detection networks perform poorly. In this paper, we propose a domain adaptation method based on adversarial learning to ensure the features extracted from a similar distribution, even when the input images belong to different domains. Considering the scarcity of images taken under certain weather conditions in the existing dataset, a semi-supervised learning framework is incorporated to enhance the detection performance through training with unlabeled images. The experiments conducted on public and private datasets show that our proposed adversarial learning technique outperforms the recent traffic scene object detection networks in all different domains. Source code and datasets are available at https://github.com/daniel851218/all-weather-vehicle-detector .