A Benchmark for Unsupervised Anomaly Detection in Multi-Agent Trajectories

Abstract

Despite the significant research efforts on trajectory prediction for automated driving, limited work exists on assessing the prediction reliability. To address this limitation we propose an approach that covers two sources of error, namely novel situations with out-of-distribution (OOD) detection and the complexity in in-distribution (ID) situations with uncertainty estimation. We introduce two modules next to an encoderdecoder network for trajectory prediction. Firstly, a Gaussian mixture model learns the probability density function of the ID encoder features during training, and then it is used to detect the OOD samples in regions of the feature space with low likelihood. Secondly, an error regression network is applied to the encoder, which learns to estimate the trajectory prediction error in supervised training. During inference, the estimated prediction error is used as the uncertainty. In our experiments, the combination of both modules outperforms the prior work in OOD detection and uncertainty estimation, on the Shifts robust trajectory prediction dataset by 2.8 % and 10.1 %, respectively. The code is publicly available 4 .

Detecting Socially Abnormal Highway Driving Behaviors via Recurrent Graph Attention Networks

“…Specifically, NetWalk [54] and TADDY [29] deal with unattributed graphs with no node or edge features; AddGraph [55] and StrGNN [7] detect anomalous edges. The most relevant work with ours is STGAE [51], where a spatiotemporal graph autoencoder is combined with kernel density estimation (KDE) to detect abnormal driving behaviors. Yet while STGAE works well in experiments with only two vehicles, the time complexity of KDE is too high for detection in large numbers of vehicles.…”

“…Compared with works [33,51] that pre-define edge weights as a function of physical distances, the attention mechanism we use has more expressive power, and can comprehensively determine the neighbor relevance based on observation information. Furthermore, the real influence of vehicles in front of and behind an ego car is asymmetric (e.g., you must slow down immediately for slowmoving cars in front of you, but not for slow cars behind you), and the attention formulation in (3) can achieve such asymmetry.…”

“…Furthermore, the real influence of vehicles in front of and behind an ego car is asymmetric (e.g., you must slow down immediately for slowmoving cars in front of you, but not for slow cars behind you), and the attention formulation in (3) can achieve such asymmetry. In comparison, when using a function of distance as edge weights [33,51], the same importance is assigned to cars that are close to a vehicle, regardless if they are in front or behind the ego vehicle.…”

“…We compare our method with simple heuristic methods as well as state-of-the-art methods for anomaly detection on trajectory data. The baselines include: i) Linear temporal interpolation (LTI) implemented by [51] that uses a linear interpolation between the first and last position of the vehicle to reconstruct the trajectory; ii) Constant Velocity Model (CVM) [37] that assumes constant speed as recorded at the first observation time step to reconstruct the trajectory; iii) Robust tensor Recovery (RTR) [20] model that captures spatial-temporal correlations via low-rank tensor decomposition, and detects sparse outliers that deviates from the normal patterns; iv) Seq2Seq model [42] that encodes and decodes the time series with two LSTM networks, and uses Mean Square Error (MSE) as the reconstruction loss; v) Spatio-temporal graph auto-encoder (STGAE) [51] that uses convolutional networks temporally and graph convolutional networks spatially, and uses bi-variate Gaussian reconstruction error to build autoencoder to derive anomaly score 3 ; vi) DSAB-biv, which is the variant of our model, that uses the same RGAT structure, but uses the bi-variate Gaussian loss as in [33,51]. Each of the baseline methods produces an anomaly score for the vehicles in each time step within a time window, and the same process is used following Eq (8) and ( 9) to calculate the vehicle and scene anomaly scores.…”

Learning Representation for Anomaly Detection of Vehicle Trajectories

“…Reconstruction-based methods assume that anomalies are not compressible and thus cannot be reconstructed from low-dimensional projections [17]. In recent works, deep generative models, such as variational autoencoder (VAE) [18]- [20], Generative Adversarial Networks [21], [22] and adversarial autoencoder [23], are introduced to perform reconstruction-based anomaly detection. One-class classification methods including one-class SVM (OC-SVM) [24], [25] and one-class neural network (OCNN) [26] are designed to learn a discriminative boundary surrounding the normal samples.…”