Animal-vehicle collision mitigation using deep learning in driver assistance systems

Abstract

The significant rise in the rate of animal-vehicle collision on Indian roads in recent years underline the imperative need for effective technological interventions that help mitigate animal-vehicle collisions in real-time. This paper proposes a deep learning approach for the implementation of the Real-Time Animal Vehicle Collision Mitigation System. The paper makes a comparative study of three deep learning based object detection frameworks: (1) Mask R-CNN; (2) You Only Look Once (YOLOv3); and (3) MobileNet-SSD. These object detectors are implemented, evaluated and compared with each other on the basis of their mean average precision (mAP) and processing times. The results show that the MobileNet-SSD architecture gives the best trade-off between the mAP and processing time and is best suited for Real-Time Animal Detection among the three approaches considered. Using the trained model of MobileNet-SSD based object detection framework, an Animal Vehicle Collision Mitigation System is implemented and shown to perform well on several real-world test scenarios.