Using Mask R-CNN and distance transform to detect elephants on railway tracks for collision prevention

Abstract

Train accidents are causing elephant deaths and habitat loss in the country. To prevent this, a real-time elephant Intrusion Detection System (IDS) has been developed that encompasses four modules viz., dataset collection and compilation, pre-processing, segmentation by Mask Region-based Convolution Neural Network (Mask R-CNN) framework, and classification by Euclidean Distance Transform (EDT). Initially, the elephant images were collected in an intricate forest environment, and the features of the elephant and the railway track were extracted using the combination of ResNet-101 and the feature pyramid network. The trained output features are segmented as an elephant and track. Then the distance between the elephant and the track from the segmented output is calculated by EDT which classifies an elephant that is on the track, off the track, or that is near the track. The proposed model obtained a better mean Average Precision ([email protected]–0.95) of 0.85 at a less testing computational time of 0.2 s.