MASK R-CNN for Pedestrian Crosswalk Detection and Instance Segmentation

Abstract

Pedestrians are the most exposed to accidents considering that the majority of motorists exclude them as road users. In this study, object detection using Mask Region-Based CNN and instance segmentation was applied to a pedestrian crosswalk. Training was done using Mask R-CNN for object detection with ResNet-101 backbone, with 0.001 learning rate and 2 images per GPU during 30 epochs of 100 batches. Based on the study, 500 pedestrian crosswalk images were gathered and selected for validation and training. 80% of the images are for the training set and for the validation set is 20%. Another 30 pedestrian crosswalk testing images were gathered for the model evaluation to verify the trained model stability and reliability. All 30 testing images had been detected and the accuracy of the detections is greater than 97%. If there is 2 or more pedestrian crosswalk in an image, it will make the color of MASK different to each detection. The summary test results verified that all gathered data was higher than 97% to be able to detect a pedestrian. With this, the proposed framework can detect pedestrian crosswalks using MASK Region-Based Convolutional Neural Network.