Refining the Efficiency of R-CNN in Pedestrian Detection

Abstract

The use of pre-trained deep convolutional neural networks for the purpose of enhancing the performance of detectors such as the region-based convolutional neural networks has demonstrated an exceptional role in the field of pedestrian detection. There have been various methods that have been investigated with the intent of improving the detection results of pedestrian detectors. In particular, most deep convolutional neural network techniques have shown advanced results in recent experimentation and studies. The primary reason for this improvement in performance based on deep neural networks is the ability of deep networks to learn substantial mid-level and high-level image features and also generalize well in complex images. In this study, an advanced deep learning technique based on R-CNN is refined and investigated for better performance in pedestrian detection on four specific pedestrian detection databases. The experiments involve the application of a deep learning feature extraction model in conjunction with the R-CNN detector. The deep learning feature extraction models employed are the AlexNet, VGG16 and the VGG19. The architecture of the R-CNN is re-modelled for enhanced performance by incorporating deep stacking networks and altering the activation function from sigmoid to Rectified Liner Unit (ReLU). The results of the experiments across the four investigated datasets provide significant findings about the performance of R-CNN detector when trained on different pre-trained deep convolutional neural networks with the application of deep stacking networks and ReLU activation function.