A Cognitive Urban Collision Avoidance Framework Based on Agents Priority Using Recurrent Neural Network

We propose a novel cognitive collision avoidance (CA) framework for autonomous driving (AD) vehicles in urban environments. In this framework, a hybrid future trajectory predictor is developed, which consists of a static agent classifier, a recurrent neural network (RNN) based trajectory predictor and a lane-based kinematic model predictor. To fuse the outputs of different predictors, an iterative multivariate Gaussian weighted algorithm is designed to drop outliers and estimate the predicted dynamic features more reliably. Subsequently, fed in with the fused results of observed agents, together with the current dynamic features and planned trajectory of the ego vehicle, an RNN-based priority prediction engine is applied to infer the priority probabilities distribution for CA decision, which indicates the likelihood that the vehicle continue driving according to its planned trajectory. By observing surrounding agents' historical ground truth trajectory and taking the road geometry constraints into consideration, the future dynamic features, priority probabilities distribution and the CA decision can be figured out at every timestamp cognitively and adaptively. The performance of this framework is evaluated on a prototype car in multiple typical USA urban scenarios, comparing with conventional CA systems which assume constant velocity and only work when observed agents follow traffic rules, our framework alleviates these limitations and achieves encouraging results in terms of the priority distribution estimation, with a frequency >20Hz, which is capable of running in real-time.