Collision Probability Estimation based on Prediction Model Adaptation

Abstract

Previous studies have developed various Advance Driver Assistance Systems (ADASs) to prevent collisions with preceding vehicles. Predicting collisions in advance and thus mitigating accidents is the core technology in ADAS; this technology is called Collision Risk Assessment (CRA). The key factor of such a system is to measure driving environments and predict vehicles’ future paths. However, this is challenging problem because of noise in the sensor and the front vehicle’s unknown future movement. This paper attempts to solve these limitations through studying a stochastic CRA. The process of predicting probabilistic CRA generates the future movements of vehicles from their current measurements. This requires a prediction model that represents the motion characteristics of vehicles mathematically. Kinematic-based prediction models are widely used due to their advantages of a simple structure and use of easy-to-observe information. However, a single kinematic model-based prediction has accuracy limitations because it is difficult to represent all complex situations using only one model. Thus, a method that switches models according to certain conditions was proposed, but this yielded the discontinuous probability of collision. This paper proposes the estimation method for collision probability with a model probability that indicates the degree of how well it matches the driving situation. The algorithm calculates the model probability and estimates the collision probability by higher weighting the model that more closely matches the current situation. The proposed method has the advantage of estimating the continuous collision probability by reflecting the driving situation in real time. This paper compared an experimental evaluation with the proposed algorithm and the single model-based method using scenarios in which various driving situations were developed.