IMM기법을 이용한 다중모델 기반 주변차량 경로예측 알고리즘

Abstract

ADAS and Autonomous Driving is a technology that helps drivers safely maneuver through traffic, and prevents accidents from happening by evading potential dangers. This technology is currently being actively studied worldwide. To detect potential dangers on road and decide actions to evade, the vehicle has to predict nearby objects’ movements correctly. Vehicle path prediction can be categorized into two large groups : physics based model and network based model. Physics based model predicts short term paths with high precision. This model doesn’t take into account the context of the environment, thereby is inappropriate for long term path prediction. Network based model receives the context of the scene as an input, so it performs better at predicting long term paths. But physical movement of the vehicle is not taken into account when generating paths, making unrealistic predictions. In this paper, path prediction algorithm considering both dynamic characteristics of vehicle and scene context is proposed for short-term, long-term prediction. Constant velocity, constant turn rate and velocity model is used in Extended Kalman Filter for interacting multiple models algorithm to predict physics based prediction, and time series data is used in LSTM for Network based prediction. Two predicted outputs are then combined by a weighting function in respect to prediction time. The proposed algorithm is verified by Argoverse opendataset, and showed enhanced results compared to individual models’ results.