Analysis of the Influence of Training Data on Road User Detection

Guindel, Carlos; Martin, David; Armingol, Jose Maria; Stiller, Christoph · 2018 · Crossref

DOI: 10.1109/icves.2018.8519510

archive: archived pipeline: cataloged verified

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

Summary

This dissertation addresses the critical challenge of reliable environmental perception for autonomous vehicles, specifically focusing on the joint detection and pose estimation of road users. The research is motivated by the necessity for automated driving systems to accurately identify potential hazards and determine their spatial orientation and location to ensure safe navigation in crowded traffic environments. The author aims to provide commercially viable solutions by leveraging deep neural networks, extending the popular Faster R-CNN framework to handle both object detection and viewpoint estimation using only appearance information, thereby enhancing robustness against environmental errors. The methodology centers on adapting the Faster R-CNN architecture for traffic scenes. The study introduces enhancements to the detection framework, including hyperparameter tuning and the integration of stereo vision data to encode depth information. A primary contribution is the extension of the network to estimate the orientation (viewpoint) of detected objects based exclusively on visual appearance. This capability is then exploited to propose two algorithms for 3D object localization: one utilizing stereo vision data and another using LiDAR data. The research also includes an automatic extrinsic calibration method for stereo-LiDAR sensor setups. All proposed methods are validated through systematic experimentation on the KITTI Vision Benchmark Suite, a widely recognized public dataset for autonomous driving research. The findings demonstrate that the proposed extensions to Faster R-CNN achieve high detection accuracy at near real-time rates. The integration of stereo vision improves detection performance by providing additional depth cues. The viewpoint estimation module successfully predicts object orientation, which significantly aids in 3D localization. The analysis of training data influence reveals specific factors affecting performance, such as the number of proposals, viewpoint bin resolution, and feature extractor architecture. The experiments confirm that the combined approach provides effective situational awareness, accurately modeling the position and orientation of potential hazards in the vehicle's surroundings. The significance of this work lies in its contribution to the development of robust perception systems for intelligent vehicles. By enabling joint detection and pose estimation from monocular or stereo cameras, the methods reduce reliance on expensive sensors like LiDAR for orientation tasks. The thesis validates the viability of implementing these solutions in real-world vehicles, offering a pathway toward safer autonomous navigation. The results underscore the importance of appearance-based orientation estimation in building comprehensive environmental models, addressing a key technical hurdle in the deployment of autonomous driving technologies.

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.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

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

• distraction detection algorithms

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Methodological Resource: tool software