Pedestrian detection in front of the ego vehicle using (stereo) camera in the urban scene: Deep versus Shallow learning approaches

Abstract

Object detection is crucial in the environment of autonomous driving and advance driver assistance systems for safely maneuvring vehicle in the urban traffic. Among the traffic participants we find pedestrians are the one who are most vulnerable and their safety is also crucial. Therefore, this work focuses on pedestrian detection in urban environment using the camera mounted on ego vehicle. The thesis aims at understanding and comparison of shallow and deep learning approaches for pedestrian detection, and two ensemble methods are proposed that combines the chosen deep and shallow method with the context-based classifier respectively. Firstly, an pre-trained deep architecture for object detection is combined with the context-based classifier. Whereas, in second method shallow approach is combined with context-based classifier. Further in the outlook of this work stereo data is used to minimize the detected false positives form the proposed ensemble deep approach. Prototyping of first proposed method is achieved using the CAFFE deep learning framework with Python interface, and the second shallow method is achieved using the well known computer vision library OpenCV with C++. The proposed method is trained, tested and evaluated on Caltech pedestrian dataset with di↵erent metric