Traffic lights perception problem is one of the key challenges for autonomous vehicle controllers in urban areas. While a number of approaches for traffic light detection have been proposed, these methods often require a prior knowledge of map and/or show high false positive rates. Recent successes suggest that deep neural networks will be widely used in self-driving cars, but current public datasets do not provide sufficient amount of labels for training such large deep neural networks. In this paper, we developed a two-step computational method that can detect traffic lights from images in a real-time manner. The first step exploits a deep neural object detection architecture to fine true traffic light candidates. In the second step, a point-based reward system is used to eliminate false traffic lights out of the candidates. To evaluate the proposed approach, we collected a human-annotated large-scale traffic lights dataset (over 60 hours). We also designed a real-world experiment with an instrumented self-driving vehicle and observed that the proposed method was able to handle false traffic lights substantially better compared with the baseline considered.