One of the main causes of traffic collisions for cyclists is incoming vehicles in blind spots. In this paper, we present a helmet-mounted camera system that can identify moving objects which are considered potentially dangerous for cyclists. We first compute the road areas and allocate the positions of vehicles in the image sequences based on the road properties, the Haar-like features, and AdaBoost cascade classifier. Then, we introduce a new algorithm based on ORB features matching to evaluate the probability of collision with the cyclist. A prompt warning sign will show up if a collision is classified as the most likely. Several experiments were conducted in crowded traffic areas under different weather conditions. The experimental results show that our algorithm is highly effective in terms of predicting the danger due to moving vehicles in blind spots and can be processed in real-time.