Haar Wavelet Based Block Autoregressive Flows for Trajectories

Abstract

Prediction of trajectories such as that of pedestrians is crucial to the performance of autonomous agents. While previous works have leveraged conditional generative models like GANs and VAEs for learning the likely future trajectories, accurately modeling the dependency structure of these multimodal distributions, particularly over long time horizons remains challenging. Normalizing flow based generative models can model complex distributions admitting exact inference. These include variants with split coupling invertible transformations that are easier to parallelize compared to their autoregressive counterparts. To this end, we introduce a novel Haar wavelet based block autoregressive model leveraging split couplings, conditioned on coarse trajectories obtained from Haar wavelet based transformations at different levels of granularity. This yields an exact inference method that models trajectories at different spatio-temporal resolutions in a hierarchical manner. We illustrate the advantages of our approach for generating diverse and accurate trajectories on two real-world datasets – Stanford Drone and Intersection Drone. Electronic supplementary material The online version of this chapter (10.1007/978-3-030-71278-5_20) contains supplementary material, which is available to authorized users.

FloMo: Tractable Motion Prediction with Normalizing Flows

“…Most existing generative models only possess some of these properties. In contrast to the flow-based prediction models proposed in concurrent works [31], [32], the flow we use is based on splines and hence is more flexible, which our results demonstrate. Furthermore, we propose a novel noise injection method that significantly stabilizes training and a data augmentation transformation that further improves our model's generalization and performance.…”

“…The larger Stanford drones dataset contains 10300 individual traffic participants, it covers roads and besides pedestrians it includes also other agent types like cyclists and vehicles. All datasets are heavily used in the motion prediction domain [16], [15], [21], [31], [2], [1].…”

“…Bhattacharyya et al [8] use a conditional Flow VAE with condition and posterior regularization to predict trajectories. In their recently published work [31], they use a block autoregressive flow based on Haar wavelets to learn distributions for motion prediction and also adapted FlowWaveNet [27] for motion prediction. Ma et al [32] recently showed how to find those trajectories sampled from affine flows that are both likely and diverse to make predictions.…”

Euro-PVI: Pedestrian Vehicle Interactions in Dense Urban Centers

“…Normalizing flows are combined with a context attention mechanism in [32] to improve expressiveness in the latent space. HBA-Flow [5] improves trajectory prediction with a Haar-wavelet based decomposition. DESIRE [23] uses a RNN refinement module over the plain cVAE setup.…”

Multiple Path Prediction for Traffic Scenes using LSTMs and Mixture Density Models

“…In (Bartoli et al, 2018) multiple cameras were used to predict the trajectory of people in crowded scenes and (Kim et al, 2017) predict the trajectory of vehicles in an occupancy grid from the point of view of an ego-vehicle. A more closely related work to this paper is presented in (Bhattacharyya et al, 2017), here they predict the future path of pedestrians using RNNs as encoder-decoders and also include the prediction of the odometry of the ego-vehicle.…”