Improved candidate generation for pedestrian detection using background modeling in connected vehicles

Abstract

Pedestrian detection is widely used in today's vehicle safety applications to avoid vehicle-pedestrian accidents. The current technology of pedestrian detection utilizes onboard sensors such as cameras, radars, and Lidars to detect pedestrians, then information is used in a safety feature like Automatic Emergency Braking (AEB). This paper proposes pedestrian detection system using vehicle connectivity, image processing and computer vision algorithms. In the proposed model, vehicles collect image frames using on-vehicle cameras, then frames are transferred to the Infrastructure database using Vehicle to Infrastructure communication (V2I). Image processing and machine learning algorithms are used to process the infrastructure images for pedestrian detection. Background modeling is used to extract the foreground regions in an image to identify regions of interest for candidate generation. This paper explains the algorithms of the infrastructure pedestrian detection system, which includes image registration, background modeling, image filtering, candidate generation, feature extraction, and classification. The paper explains the MATLAB implementation of the algorithm with a road-collected dataset and provides analysis for the detection results with respect to detection accuracy and runtime. The algorithm implementation results show an improvement in the detection performance and algorithm runtime. © 2020 by the authors.