Vehicle taillight recognition is an important application for automated driving, especially for intent prediction of ado vehicles and trajectory planning of the ego vehicle. In this work, we propose an end-to-end deep learning framework to recognize taillights, i.e. rear turn and brake signals, from a sequence of images. The proposed method starts with a Convolutional Neural Network (CNN) to extract spatial features, and then applies a Long Short-Term Memory network (LSTM) to learn temporal dependencies. Furthermore, we integrate attention models in both spatial and temporal domains, where the attention models learn to selectively focus on both spatial and temporal features. Our method is able to outperform the state of the art in terms of accuracy on the UC Merced Vehicle Rear Signal Dataset, demonstrating the effectiveness of attention models for vehicle taillight recognition.

No abstract

The difficulties in analyzing large and extensive systems necessitate the use of efficient machine-learning tools to identify unknown system anomalies in order to avoid critical problems and ensure high reliability. Given that data logged by a system include unknown anomalies, anomaly identification models aim to simultaneously identify the time of occurrence and the features that contributed to these anomalies. To maximize accuracy, it is important to utilize the data as well as the domain knowledge of the system. However, it is difficult for a system analyst to possess not only machine-learning capabilities but also domain knowledge to incorporate into the model. In this paper, we propose a new anomaly identification framework capable of utilizing feedback based on domain knowledge without requiring any machine-learning capabilities. We also propose a novel method, the so-called rank ensemble method, to improve the accuracy of anomaly identification with erroneous feedback, that is, feedback that in- cludes incorrect information. Our method enables erroneous information to be adaptively ignored by assuming consistency between the data and the user feedback. An intensive parameter study using benchmark datasets and a case study with real vehicle data demonstrate the applicability of our framework.

In this paper, we propose a novel framework to help human operators- who are domain experts but not necessarily familiar with statistics- analyze a complex system and find unknown changes and causes. Despite the prevalence, researchers have rarely tackled this problem. Our framework focuses on the representation and explanation of changes occurring between two datasets, specifically the normal data and data with the observed changes. We employ two-dimensional scatter plots which can provide comprehensive representation without requiring statistical knowledge. This helps a human operator to intuitively understand the change and the cause. An analysis to find two-attribute pairs whose scatter plots well explain the change does not require high computational complexity owing to the novel characteristic function-based approach. Although a hyper-parameter needs to be determined, our analysis introduces a novel appropriate prior distribution to determine the proper hyper-parameter automatically. The experimental results show that our method presents the change and the cause with the same accuracy as that of the state-of-the-art kernel hypothesis testing approaches, while reducing the computational costs by almost 99% at the maximum for all popular benchmark datasets. The experiment using real vehicle driving data demonstrates the practicality of our framework.

No abstract