Top-down based saliency model in traffic driving environment

Traffic driving environment is a complex and dynamically changing scene. During driving, drivers always focus their attention on the most important and saliency areas or targets. Traffic saliency detection is an important application area of computer vision, which could be used to support autonomous driving, traffic sign detection, driving training, and car collision warning, etc. At present, most saliency approaches are based on bottom-up computation which does not consider the top-down control and cannot match the actual traffic saliency in drivers´ eyes. In this paper, by carefully analyzing the eye tracking data of 40 subjects who were non-drivers and drivers when viewing 100 traffic images, we found that the drivers´ attention was mostly concentrated on the front of road. We proposed that the vanishing point of road can be regarded as top-down guidance in the traffic saliency model. Subsequently, we gave the framework of a bottom- up and top-down combined traffic saliency model, and the results showed that our method can effectively simulate the attentive areas in driving environment.