Anomaly detection in spatiotemporal data is a challenging problem encountered in a variety of applications including hyperspectral imaging, video surveillance, urban traffic monitoring and environmental monitoring. Existing anomaly detection methods are mostly suited for dealing with point anomalies in sequence data. These methods cannot deal with temporal and spatial dependencies that arise in spatiotemporal data. In recent years, tensor based anomaly detection methods have been proposed to deal with the multi-way structure inherent to these data. Most of the tensor based methods are supervised or semi-supervised, which rely on training models based on labeled cases and assume relatively stable patterns in normal cases. In this paper, we introduce an unsupervised tensor based anomaly detection method for spatiotemporal urban traffic data. The proposed method assumes that the anomalies are sparse and temporally continuous, i.e. anomalies appear as spatially contiguous groups of locations that show anomalous values consistently for a short duration of time. Moreover, we preserve the local geometric structure of the data through graph regularization across each mode. The proposed framework, Graph Regularized Low-rank plus Temporally Smooth Sparse decomposition (GLOSS), is formulated as an optimization problem and solved using ADMM. The resulting algorithm is shown to converge and be robust against missing data and noise. The proposed framework is evaluated on both synthetic and real spatiotemporal urban traffic data and compared with baseline methods.