Data Transformer for Anomalous Trajectory Detection

Abstract

Anomaly detection is an important task in many traffic applications. Methods based on convolutional neural networks reach state-of-the-art accuracy; however, they typically rely on supervised training with large labeled data and the trained network is only applicable to the intersection that the training data are collected from. Considering that anomaly data are generally hard to obtain, we present data transformation methods for converting data obtained from one intersection to other intersections to mitigate the effort of training data collection. We demonstrate our methods on the task of anomalous trajectory detection and leverage an unsupervised method that require only normal trajectories for network training. We proposed a general model and a universal model for our transformation methods. The general model focuses on saving data collection effort; while the universal model aims at training a universal network for being used by other intersections. We evaluated our methods on the dataset with trajectories collected from GTA V virtual world. The experimental results show that with significant reduction in data collecting and network training efforts, our methods still can achieve state-of-the-art accuracy for anomalous trajectory detection.

Tasks

Reproductions