|
Abstract:
|
This paper presents a novel time-series classification approach for imbalanced data based on kinematic signatures. The method consists of two components: automatic feature extraction based on CNN-GRU and classification based on XGBoost. For feature extraction, CNN can precisely catch features from local time slices while GRU focuses on the time dependence among those slices to get the global information for a given segment, and apply weighted categorical cross-entropy loss to alternate the relative importance by allocating more weight to the less represented crash. For classification, we utilize a XGBoost classifier instead of the original full connection layer to avoid high class-weight risk in the neural network and enhance the robustness of the model. Experiments show that in 3-class classification(crash, near-crash, normal driving ), the accuracy for the overall model is 97.5%, the precision and recall for crashes are 82.5% and 72.14% respectively, which is substantially better than benchmarks models. Furthermore, the recall of most severe crashes is close to 100\% while the recall of minor crashes and low-risk tire strikes is more than 82%.
|
