Pedestrian Detection Using MobileNetV2 Based Mask R-CNN

Abstract

AbstractPedestrian detection system is one of the recent technological innovations to save human lives on the roadways. According to WHO, road accidents highly contribute to the increasing mortality. These traffic accidents can be avoided by utilizing an autonomous vehicle equipped with AI algorithms to identify the pedestrians efficiently. Advances in computer vision and deep learning techniques open up new research possibilities. This research study has introduced novel algorithms to combine two of the most efficient deep learning models. It is well-known that the Mask R-CNN is the most efficient deep learning model used for performing object detection in two stages. This process leverages high accuracy but it also include limitations such as low detection speed and high computational cost. To resolve this problem a lighter version of Mask R-CNN with MobileNetV2 architecture has been developed. In order to make Mask R-CNN light, some modifications have been done in Region Proposal Network (RPN) like a Convolution Operation is replaced by the Depthwise Separable Convolution Operation. To further speed up the process, MobileNetV2 architecture is used in the place of ResNet-101. MobileNetV2 uses inverted residual block and linear bottleneck to generate less number of parameters and reduced network computation to save time and speed up the process. The main goal of the proposed model is to detect the pedestrian with high accuracy and at high speed by consuming less computational cost without compromising robustness of the system. Further, the experimentation has been carried out with INRIA dataset and recorded a 98.9% detection rate, 0.87 mAP and 0.85 mIoU, which is far better than the standard Mask R-CNN with ResNet-101 architecture. The proposed model's weight is also 65% lighter than the conventional model, allowing it to operate faster and spend less time in interpretation. The performance of ResNet-101 based Mask R-CNN and MobileNetV2 based Mask R-CNN were compared in this study. This model can be expanded in future to yield more accurate findings and improve the ability to deal with emerging challenges in smart vehicles. KeywordsMask R-CNNMobileNetV2Pedestrian detectionAutonomous vehicleDeep learningSegmentation