A deep learning based autonomous distance estimation and tracking of multiple objects for improvement in safety and security in railways

Authors: Muhammad Abdul Haseeb, Danijela Ristic-Durrant, Axel Graser

Abstract: In this paper, a novel approach for multiple object tracking and distance estimation from an on-board monocular camera, aiming at improvements in the safety and security of railways, is presented. The approach is based on deep learning architecture using a deep Convolutional Neural Network (CNN) object detection followed by a multi hidden-layer Gated Recurrent Neural Network (RNN) referred as DisNet-RNN Tracker, which consists of two sub-networks for distance estimation and bounding box prediction respectively. The DisNet-RNN Tracker learns and estimates the distance between the detected object and the camera sensor, and predicts the object bounding box based on sequential input from previous and current detection. The presented DisNet-RNN Tracker tracks multiple objects in case where object detection module fails to detect object. The presented method is evaluated on the real-world railway dataset recorded with the on-board Obstacle Detection System developed within a H2020 Shift2Rail project SMART Smart Automation of Rail Transport. The presented work has potential to benefit other applications where reliable object detection, tracking and long-range distance estimation is needed such as autonomous cars, transportation and public security.

---

---