Occluded pedestrian detection by pre-detecting potentially dangerous regions

Abstract

A pedestrian walking out suddenly from behind a parked vehicle is a dangerous occurrence for drivers and pedestrians alike. Through experience, human drivers have learnt to anticipate and avoid accidents that can arise from such events. Unlike human drivers, machines struggle to detect occluded pedestrians in a quick manner, as desired in a sudden crossing pedestrian scenario. This project focuses on pre-detecting the regions in which a driver anticipates a suddenly crossing pedestrian, as well as determining if the detected region is ‘Safe’ or ‘Dangerous’ for an assisted, or autonomous, driving scenario. This project proposes a new method for detecting sudden pedestrian crossing events. The method uses video or frames captured from the perspective of a vehicle to detect the gaps between parked vehicles, referred to as blind spots. These blind spots are cropped images from the scene which contain features from the edge of the parked vehicle, from the background information, and also features from any pedestrian that may be emerging from behind the parked vehicle. A dataset had to be collected in order to obtain enough instances of people emerging suddenly from behind a parked vehicle. This dataset was used to train a model, consisting of a pre-trained VGG19 backbone model integrated with a binary classifier, by transfer learning. The purpose of this integrated model is to take the blind spot images as input, extract their relevant features, and finally classify the blind spot as either ‘Safe’ or ‘Dangerous’. A ‘Dangerous’ blind spot indicates that there is a pedestrian emerging from within that blind spot, as opposed to the ‘Safe’ blind spot. The method proposed produced an average accuracy of 93.84% at detecting ‘Dangerous’ blind spots and 89.39% detecting ‘Safe’ blind spots, with an execution time of less than 250ms. The proposed method, when compared to other means of detecting occluded pedestrians, obtained a higher accuracy. Therefore, the method produced accurate results with an execution time that is comparable to a human reaction time of 250ms and, hence, quick enough for the application of sudden pedestrian detection.