Few-shot recognition of lane-changing intention in connected vehicle environment: A meta-transformer approach

This paper proposes a Meta-Transformer model for few-shot accurate lane-changing intention recognition in connected vehicle environment. Given the velocity, acceleration, distance headway, time headway and time to collision of a vehicle at past time steps, a transformer model is developed to recognize the lane changing intention of the vehicle at the current time step. The multi-head attention layer of the Transformer model is capable of effectively extracting features from the multivariate time serie input. This enhances the accuracy of lane-changing intention recognition. The Transformer model is combined with a model -agnostic meta-learning (MAML) algorithm to be able to perform well with small amount of training data. The meta-learning consists of meta-training and meta-testing. During the meta-training, a task is sampled from the distribution over tasks. The tasks include recognition of intention of lane-keeping, lane-changing to left lane, and lane-changing to right lane. The model is trained with few samples and feedback from the loss function defined on the tasks. The loss function is defined as the categorical cross entropy of the errors between the labels and the results generated by the proposed model. During the meta-testing, the model is tested on new data from the distribution of the transition from a feature and the label to another feature. The model is trained by minimizing the test error. Experimental results show that the proposed model can achieve 8.7 percentage points higher accuracy than a transformer model.