The 11th Iranian and the first International Conference on Machine Vision and Image Processing

A New Approach Toward Pedestrian Detection Based on a Mixture of Region Proposal and Semantic Segmentation Deep Convolution Neural Networks

Mohammad Ali Alavianmehr, Mohammad Sadegh helfroush, Habibollah Danyali, Ashkan Tashk
The 11th Iranian and the first International Conference on Machine Vision and Image Processing (MVIP 2020)


Pedestrian detection is a challenging subject for almost all systems and technologies related to the issues such as traffic flow control, intelligent video surveillance, personal identification, people tracking, advanced driver assistance systems (ADASs), robotics, and most notably pedestrian protection systems (PPSs). In intelligent video surveillance, it provides essential information for object counting, event recognition, and semantic understanding of videos. To make a headway on pedestrian detection, plenty of classification approaches as well as deformation models have been employed. Although considerable endeavors have been made, it is still considered as a daunting challenge which is mainly the result of tiny and occluded appearances, cluttered backgrounds, and adverse illumination conditions. Pedestrian detection based on Convolutional Neural Network (CNN) has achieved substantial progress. The leading process of CNN based methods can be broken up to two steps: proposal extraction and CNN classification. First, by virtue of the traditional pedestrian detection algorithm the candidate proposals are extracted. Then, by means of the CNN model, these proposals are divided into pedestrian or non-pedestrian regions. In other words, each layer in CNN consists of distinct discriminative qualities, which can be utilized for learning the classifier. In this paper, a mixture of region proposal and semantic segmenting deep neural networks are employed for detection of pedestrians from diverse databases like CUHK-Occlusion and PennFudanPed. The proposed strategy is a combina38 ■ | Book of Abstracts | MVIP 2020 | University of Tehran tion of modern RPN CNNs like YOLO and semantic segmenting networks like fully Convolutional Networks (FCNs) specifically the ones with architectures analogous to those of U-Nets. The simulations and implementation results demonstrate that the proposed method can detect pedestrians and individuals from well-known databases with acceptable a

Keywords: Deep Learning CNN; Occlusion Handling; Pedestrian Object Detection; Region Proposal Networks (RPN); Semantic Segmentation

© 2017-2021 ISMVIP All Rights Reserved