A detection and alert system for animal road crossings in electric vehicles using computer vision

Abstract

The purpose of detection and alert systems in electric vehicles is to inform the driver of impending collisions with other objects along the path which the vehicle is moving. Electric vehicles use motors and batteries to provide the propulsion force for the car. Integrating embedded systems with such is much easier than with internal combustion engines. Animal road crossings are predominant in Uganda because urban farming is widely practiced and most road networks traverse game parks/reserves. In this project , we developed a low cost and efficient detection and alert system based on a YOLOv4 object detector and a monocular camera for forward collision detection. We chose a single object (elephants) to be our case study and test for this project. We trained the object detector on an African elephant dataset and an elephant sign post dataset (open).We further implemented a distance ,speed and direction estimator in python based on the triangular similarity using a single reference image with known distances to determine the distances of the new objects detected. Our results for the object detector were based on the @mAP scores achieving accuracies of 93.56% and 93.4% for the elephant dataset and elephant signpost dataset. The distance estimates were accurate for fast moving frames and had a range of 20m from the camera. We were able to demonstrate implementing a detection and alert system with no additional sensors making it relatively cheaper and compatible with electric vehicle designs in Uganda.