Egocentric Vision-based Future Vehicle Localization for Intelligent\n Driving Assistance Systems

Yao, Yu; Xu, Mingze; Choi, Chiho; Crandall, David; Atkins, Ella; Dariush, Behzad · 2018 · arXiv (Cornell University)

DOI: 10.48550/arxiv.1809.07408

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of predicting the future location and scale of nearby vehicles from an egocentric (first-person) camera view, a task critical for Advanced Driver Assistance Systems (ADAS) and autonomous driving. While existing trajectory prediction methods often rely on bird’s-eye views or static cameras, this work focuses on the more practical but complex egocentric perspective, where object distance is inferred indirectly through scale and appearance changes. The authors specifically target challenging urban intersection scenarios, where vehicle trajectories are diverse and dynamic, unlike the simpler highway scenarios common in prior datasets. To solve this problem, the authors propose a multi-stream Recurrent Neural Network Encoder-Decoder (RNN-ED) architecture. The model utilizes two parallel encoders to process temporal information: one stream encodes past bounding box trajectories to capture location and scale, while the other encodes dense optical flow to capture pixel-level motion and appearance changes. These streams are fused to initialize a decoder, which predicts future bounding boxes. Crucially, the decoder also incorporates future ego-motion (planned vehicle movement) as an input, allowing the model to account for apparent motion caused by the ego-vehicle’s own actions. The authors introduce the Honda Egocentric View-Intersection (HEV-I) dataset, comprising 230 videos collected at intersections, to evaluate performance alongside the existing KITTI dataset. Experimental results demonstrate that the proposed RNN-ED model significantly outperforms baseline methods, including linear regression, constant acceleration models, and state-of-the-art Conv1D approaches. On the HEV-I dataset, the best model (RNN-ED-XOE) achieved a Final Displacement Error (FDE) of 24.92 across all cases, improving upon the Conv1D baseline by approximately 15%. Ablation studies confirmed that incorporating dense optical flow and future ego-motion substantially reduces prediction errors, particularly in challenging cases involving complex maneuvers. The model also achieved state-of-the-art results on the KITTI dataset, with an FDE of 37.11 compared to 44.13 for the Conv1D baseline. Qualitative analysis showed that the RNN-ED architecture provides superior temporal modeling compared to convolution-deconvolution models, which generate trajectories in a single shot. The significance of this work lies in its novel approach to egocentric future vehicle localization, providing a robust framework for safety-critical driving applications. By explicitly modeling both object appearance/motion via optical flow and ego-vehicle planning, the method addresses key limitations of previous egocentric prediction techniques. The release of the HEV-I dataset further contributes to the field by providing a resource for training and evaluating models in complex, interaction-heavy driving environments. The authors conclude that while the current model performs well, future work should incorporate broader scene context, such as traffic signs and vehicle-to-vehicle interactions, to handle failure cases like occlusions and uneven road surfaces.

Key finding

A multi-stream recurrent neural network that incorporates dense optical flow and future ego-motion cues significantly outperforms baseline methods in predicting future vehicle locations and scales from egocentric video feeds.

Methodology

simulation_modeling

Sample size: 2477

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via author_sweep_intake on 2026-05-27.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• anticipation