Factors Influencing The Detectability Of Pedestrians In Urban Environments

Driver assistance systems based on computer vision modules aim to provide useful information for the driving task to its user. One critical task in such scenarios is avoiding dangerous encounters between cars and vehicles. Classical computer vision systems aim only at finding all pedestrians. We propose that in order to provide the maximally useful information to the driver, it is also necessary to know the probability that the driver will see the pedestrian. This way the system is able to direct and modulate the attention of the driver towards pedestrians that he might not have noticed. Methods: We performed an experiment with 10 subjects. We showed images of urban environments for 120 ms followed by a noise mask. Afterwards, subjects had to indicate positions where they saw a pedestrian. We used the MIT StreetScenes database [1] which contains 3547 photos with hand-labeled pedestrian positions. Each participant was shown a total of 557 images in a random order. 142 images without pedestrians, 245 contained one single pedestrians and the rest contained two or more pedestrians. Results: We counted mouse clicks within a 100 pixel radius of the center of a pedestrian as hits. The average hit rate was 69. We evaluated how well a classifier can predict the detectability of a pedestrian based on several features such as: compositional features (position and size of the pedestrian), image features (color histograms, contrast and histogram of oriented gradients descriptors of the pedestrian as well as the decision value of a support vector machine trained on a pedestrian classification task) and context features (difference in mean, standard deviation and color histograms between pedestrian and background and distance to other pedestrians in the image)