HOPE: Hierarchical Spatial-temporal Network for Occupancy Flow Prediction

Abstract

In this report, we introduce our solution to the Occupancy and Flow Prediction challenge in the Waymo Open Dataset Challenges at CVPR 2022. We have developed a novel hierarchical spatial-temporal network featured with spatial-temporal encoders, a multi-scale aggregator enriched with latent variables, and a recursive hierarchical 3D decoder. We use multiple losses including focal loss and modified flow trace loss to efficiently guide the training process. Our method achieves a Flow-Grounded Occupancy AUC of 0.8389 and outperforms all the other teams on the leaderboard.

A Review of Deep Learning-Based Vehicle Motion Prediction for Autonomous Driving

“…Refs. [29,106,107] jointly predict the occupancy grid map and occupancy flow field of the region of interest based on the model's understanding of the scene. Mahjourian et al [29] use a similar input characterization and encoding process as [42] but apply a feature pyramid network (FPN) [100] to fuse multi-scale features during decoding.…”

“…Liu et al [106] consider the extraction of historical motion agent features and the global scene feature and then use a Swin-transformer [108] to consider the fusion of objectlevel features and global scene-level features. Hu et al [107] define a hierarchical spatialtemporal network with multi-scale feature fusion to encode scene feature maps in both spatial and temporal dimensions.…”

Hierarchical speed planning and energy management for autonomous plug-in hybrid electric vehicle in vehicle-following environment

“…While achieving agent-wise accuracy, MATP introduces exponential computations and trajectory-wise inconsistency. Dense predictions directly estimate the future distribution of agents jointly from ego-centered occupancy [9], [20], [21]. A notable issue is the loss of agent-wise tractability.…”