Structural Attention-Based Recurrent Variational Autoencoder for Highway Vehicle Anomaly Detection

提示：文章写完后，目录可以自动生成，如何生成可参考右边的帮助文档

---

论文链接

https://arxiv.org/abs/2301.03634

代码链接

https://gitlab.engr.illinois.edu/hubris/highway-anomaly-detection

keywords

Anomaly Detection; Autonomous Vehicles; Unsupervised Learning; Human Behavior Modeling

Abstract

We propose a novel unsupervised framework for highway anomaly de-
tection named Structural Attention-Based Recurrent VAE (SABeR-
VAE), == which explicitly uses the structure of the environment to aid
anomaly identification.==
vehicle self-attention module to learn the relations among vehicles on a road
and a separate lane-vehicle attention module to model the importance of
permissible lanes to aid in trajectory prediction
Conditioned on the attention modules’ outputs, a recurrent encoder-decoder architec-
ture with a stochastic Koopman operator-propagated latent space predicts the next states of vehicles.
modeling environmental factors is essential to detecting a diverse set of anomalies in deployment.
such interaction-aware methods still ignore the effect of road structures on vehicle
behaviors, and thus can miss abnormal scenarios like wrong-way driving trajectories that appear normal when environmental context is overlooked.

Introduction

Specifically, a neural network, which generally follows an == encoder-decoder architecture== for trajectory reconstruction or prediction, learns an underlying distribution of normal vehicle trajectories in the latent space.
An anomaly is then detected whenever the trajectory is out ofdistribution and produces a large reconstruction or prediction error.
To ensure interpretability, we use variational autoencoder (VAE) to cluster useful features from similar behaviors together in a continuous and stochastic latent space
Specifically, we treat a highway scenario as a structured interaction graph where nodes represent vehicles and lane positions, and edges connect nearby vehicles, and permissible lanes.

Related Works

Exploiting Map Information

Variational Autoencoders for Sequences

Variational autoencoders (VAE) have been applied to sequential data combined with recurrent neural networks (RNN)
To bridge the gap between complex human behaviors and the
structured environment, and overcome the hurdles of the temporal propagation in simplistic RNNs, we propose the use of a lane-conditioned Koopman Operator to model the temporal relations in the latent space. We were specifically inspired to use the Koopman operator to propagate the latent space due to its capability to model the dynamics ofcomplex, non-linear data, including fluid dynamics, battery properties, and control tasks

Anomaly Detection

In this work, we explicitly model both vehicle-to-vehicle interactions and lane-to-vehicle interactions to boost performance, and use an interpretable variational
architecture to learn a continuous distribution over behaviors.

Methodology

Problem Formulation

Architecture

Vehicle-Vehicle Self-Attention Network

Our goal is to learn a representation of spatial interactions among vehicles.
we encode the positions of vehicles on the road at each time with scaled dot-product multi-head self-attention, which allows each head to learn different features of the data


将每辆车的位移 $X_t$ 经过一个MLP得到 $Q_t$ --> 将第$i$ 辆车与相邻其他车的位移 $R_t$ 经过两个不同的MLP得到 K,V --> 经过 self-attention layer 得到 fianl encoding --> learned a representation of spatial interactions among vehicles.

Lane-Vehicle Attention Network

mask out impermissible lane nodes.

Recurrent Encoder

GRU参考文章：https://zhuanlan.zhihu.com/p/32481747

Latent Propagation withKoopmanOperator

KoopmanOperator ：propagate the latent distributions in time to predict the future states ofvehicles.
The Koopman operator is responsible for temporal reasoning (modeling vehicle state dynamics), while the preceding attention modules take charge of spatial reasoning.

The Decoder Network

Training and Evaluation

End-to-End Training

training objective:minimize the current reconstruction loss and one-step future prediction loss of the model.

Anomaly Detection Evaluation

EXPERIMENTAL SETUP AND RESULTS

MAAD Dataset and Augmentation

Baseline Methods

…

Latent Space Interpretation

SABeR-VAE is a variational model with a continuous latent space such that observations with similar learned characteristics are clustered closer together in the latent space.