An improved Faster RCNN for Pedestrian Detection

Abstract

Pedestrian detection plays a pivotal role in applications such as robotics, automated driving, assistive living, and surveillance. The problem of pedestrian detection, although approached by many computer vision researchers is far from solved. The scale, pose, occlusion, illumination, and many such factors affect the performance of the methods. In this work, a modification of the most commonly used deep convolutional neural network model ResNet18 is proposed. The modified CNN structure forms the base of the Faster RCNN model utilized to predict the locations of pedestrians in the image. The proposed method has been improved in terms of the feature map extraction of the image. To evaluate the proposed method, two benchmark datasets INRIA Pedestrian and PASCAL VOC 2012 are considered. The performance metrics used for evaluation are Detection Error Trade-off and Precision-Recall Curve. A statistical analysis is also conducted. The proposed method is compared against state-of-the-art detection methods.