Investigating the suitability of COSFIRE filters for pedestrian recognition

Abstract

Pedestrian recognition is a widely researched area due to its application in pedestrian detection systems as in the cases of surveillance in crowded places such as train stations or airports, automotive safety and assisted technology to the visually impaired. However, the recognition of pedestrians is not an easy task. Certain factors such as occlusion, light and appearance of pedestrians may hinder any pedestrian recognition system. Throughout the years, multiple state-of-the-art systems have been developed to cater for such issues such as Histogram of Oriented Gradients (HOG), Scale-Invariant Feature Transform (SIFT) and Speeded-Up Robust Features (SURF). These systems typically employ classification techniques such as SVM to determine whether any particular scene contains any pedestrian or not. However, some aspects of these systems still fail to handle occlusion and thus misclassification occurs. The system proposed makes use of COSFIRE filters to solve the problem of pedestrian recognition. By analyzing a given prototype of interest, COSFIRE filters are con figured and trained in an automatic process. Since classification techniques vary in speed, we propose the use of two classification techniques, namely Support Vector Machines (SVM) and Random Forests to evaluate COSFIRE filters' effectiveness in relation to pedestrian recognition. The effectiveness of this approach is demonstrated by using three well-established datasets, namely INRIA Person dataset, Daimler Mono Pedestrian Dataset and TUDBrussels Pedestrian Dataset, which all vary in size and type of images they contain. HOG descriptors are also computed to better evaluate the approach using COSFIRE filters. The highest accuracy rate achieved using COSFIRE is that of 98.67% on the INRIA Person dataset and 93.55 % on the same dataset using HOG, both times using SVM. Tests are also performed using L2-normalized data, which is applied on the tiles of the spatial pyramid, as well as tests involving cropped and uncropped images. It is concluded that unnormalized descriptors outperform normalized descriptors when the data in question is uncropped. When data is cropped, there is no significant difference if data is normalized or not. It can also be deduced that COSFIRE outperforms HOG on the INRIA and Daimler datasets. Moreover, our proposed approach outperforms other works in this field such as the 90% accuracy rate obtained by Gavrila et al. in their survey on pedestrian classification.