A Novel Multikernel Convolutional Neural Network Classification For Pedestrian Detection In Crowded Scenes

Abstract

Pedestrian detection is a basic challenge in image processing, and there are many practical applications in the fields of robotics, video surveillance, autonomous driving, and vehicle safety. However, this is still a daunting problem due to the huge differences in lighting, clothing, color, size, and posture. A pedestrian detection method based on MKCNN (Multi-kernel Convolutional Neural Network) classifier is implemented in this paper. High accuracy is achieved by automatically optimizing the feature representation of neural network recognition and regularization problems. We evaluated the proposed method in a complex database that includes pedestrians in an urban environment without being restricted by posture, motion, background, and lighting. Experiments using the pedestrian data set of the California Institute of Technology show that the use of MKCNN can improve detection accuracy. Using the CNN model, our pedestrian detection method provides a higher detection performance for the Caltech dataset.