Importance Rank-Learning of Objects in Urban Scenes for Assisting Visually Impaired People

This paper examines an importance rank learning method of objects in urban scenes for assisting visually impaired people. Object detection methods have been used to assist visually impaired people in identifying obstacles in urban scenes, such as cars and trees. However, these existing methods are not dedicated to predicting which obstacle is important. Thus, we propose a method that estimates the importance of objects and warns them to users in order of importance ranking. We introduce a neural network-based ranking estimation method to predict the importance ranking of objects. In particular, our method uses optical flow from the previous frame and region data of detected objects as input. It helps to consider states of moving objects (e.g., cars, motorbikes, people) in a scene. Experimental results show that our model outperforms three other baselines qualitatively and quantitatively. Furthermore, our method was highly evaluated than the baseline methods by qualified caregivers of the visually impaired people.