An Ensemble Learning-Online Semi-Supervised Approach for Vehicle Behavior Recognition

In autonomous vehicles, recognizing different maneuvering behaviors of surrounding vehicles is crucial to reduce traffic risks and achieve safe path planning. Conventional vehicle behavior recognition methods adopt mainly supervised learning methods and assume that many sample labels are available. However, manual sample labeling is often time-consuming and laborious. Also, onboard sensors collecting surrounding vehicle movement information in data streams often cannot process information in real-time. To tackle these problems, we propose a semi-supervised approach using K-nearest neighbor- (K-NN)-based ensemble learning to classify the maneuvering behaviors of surrounding vehicles. The framework is divided into three parts: initial model training, online classification, and online model updating. First, k-means clustering of the maneuvering behavior is performed, cluster features are calculated, and a set of micro-clusters is obtained to establish the initial model. Second, the ensemble K-NN-based learning method is used to classify the incoming instances. Finally, the model is updated online using error-driven representative learning and an exponential decay function. Typical lane-changing and turning maneuvers are used as representatives to verify the performance of the proposed method. The data are provided by a next-generation simulation project. The results show that the proposed model achieves highest average recognition accuracy compared with other benchmark methods for the lane-changing and turning maneuvers shortly after the maneuver begins, even for a small sample size.