۱٫ Introduction

۲٫ Proposed Methodology

۳٫ Experimental Outcome and Discussion

۴٫ Conclusion

۵٫ References

This paper presents, feature extraction techniques such as center symmetric local binary pattern (CSLBP), extended CSLBP (XCSLBP), speeded-up robust feature (SURF) with 64 and 128 feature descriptors and histograms of oriented gradients (HOG) applied on a set of images from INRIA person database, to detect pedestrians. About fifteen feature sets created using different combinations of the aforementioned methods are compared using two detectors, random forest (RF) and support vector machine (SVM). Performance validation is done based on the accuracy, precision, recall and space required for storing feature vectors. Experimental results have shown that CSLBP and the novel XCSLBP+CSLBP feature sets yield 100% accuracy, when used with RF classifier, whereas, the novel SURF-128+XCSLBP combination and SVM linear classifier gave 99.2% accuracy in detecting pedestrians.

Introduction

Pedestrian detection is the most challenging intelligent system applications that would be helpful in advanced driver-assistance systems (ADAS). This need to detect pedestrians is quenched with various recently proposed feature extraction and classification techniques, which can be used in the separating pedestrian images from nonpedestrian images. In the past, researchers have proposed several approaches for pedestrian detection and object detection in general. Histograms of oriented gradients (HOG)1 is one such feature extraction method, when applied in combination with linear support vector machine (SVM) on MIT pedestrian dataset close to perfect results were obtained. Furthermore, the performance of HOG was proved better using another dataset named ‘INRIA’, consisting of human images with different backgrounds and pose variations. The progress of pedestrian detection techniques in a span of ten years was studied in detail2 where HOG was the most widely used technique in combination with other methods.