A Framework for Lane Prediction Based on Vehicle Detection and Tracking

Abstract

Abstract In this paper, we propose a novel, real-time, coupled framework for vehicle detection, unsupervised learning based vehicle tracking and lane prediction by an outside looking camera mounted on the dashboard of a vehicle. We detect the vehicle using YOLOv3 (You only look once) object detection method [1] and achieve \(98.7\%\) detection accuracy on an average. Then, we apply incremental clustering across frames for tracking the vehicles on-the-fly. Moreover, we detect and track lanes using unsupervised learning based algorithm [2], wherever the lanes are visible. Many-a-time, the lane markings are not visible due to wearing-off of the markings or occlusion caused by the vehicles on the road. In our proposed framework, we use the information of vehicle tracks and detected lane markings for predicting lanes where the lanes are not present or not visible on the road. Our framework detects and tracks the vehicles accurately in each frame and successfully predicts the lanes even in challenging scenarios such as, in the presence of occlusion, illumination variation, etc.KeywordsVehicle detectionVehicle trackingLane estimationDeep learningIntelligent vehiclesLane prediction.