ConvMLP for Driving Behavior Detection from Naturalistic Driving Data

Because of the complexity and volatility of the traffic environment, driving behaviors detection in naturalistic driving is still a significantly hard problem. In this paper, a novel attention-free model, Convolutional MLP (ConvMLP) is proposed for driving behavior detection using video frames and multivariate vehicle signals. By introducing convolutions into the architecture through two aspects, i.e., a hierarchy of MLP stages containing convolutional token embeddings, and ConvMLP modules to model local and global dependencies from multi-modal data simultaneously, MLP-based model can be improved in both performance and efficiency. Specifically, driving context that includes driving scene features and vehicle dynamics features can be generated, and a Long Short-Term Memory (LSTM) neural network is designed to learn the heterogeneous features and detect driving behaviors. Moreover, the generalization ability of ConvMLP is investigated by a customized Spatial Ensemble (SE) student-teacher framework on unlabeled data. The experiment results on a self-collected driving data set show that the proposed ConvMLP with a smaller model size performs on par with or superior than other state-of-the-art models.