| Résumé | Among state-of-the-art vehicle maneuver classification algorithms, Hidden Markov Models are commonly applied for predicting maneuver probability. To generate a model, a sufficient number of labeled samples is necessary for training. However, annotations that contain information such as class and bounding box are not always available. Manually labeling data is tedious, inaccurate, and time consuming, especially when the dataset is extremely large. Besides, there exists lots of redundant data that negatively influences model training. In this paper, we explore the possibility of using only a few manually labeled samples to train a Convolutional Neural Network (CNN) model for vehicle maneuver classification with relatively high accuracy. We define three maneuver classification groups: motion, velocity, and turning. Each group has subclasses, reflecting different aspects of the vehicle movement. Based on the defined maneuver classes, we design a simple CNN model to distinguish vehicle maneuvers. We also propose a learning strategy that requires only a few samples for training, while maintaining high recognition precision. Comprehensive experiments were performed to demonstrate the capability of our model and performance of our training strategy. In result, our model achieves an overall of 93.48% precision for maneuver recognition with only 3.5% Controller Area Network (CAN) bus data for training. |
|---|
