FENet: Focusing Enhanced Network for Lane Detection

TL;DR

This paper tackles lane detection for autonomous driving by leveraging human-like visual focus on distant road regions. It proposes FENet, a framework that combines Focusing Sampling, Partial Field of View evaluation, coordinate-augmented feature pyramids, and a Directional IoU loss to improve distant and curved lane localization. Two variants, FENetV1 and FENetV2, trade off perspective-aware context versus regression-focused distance and direction modeling, with V2 showing stronger real-world applicability due to improved distant-lane regression. Experiments on CULane and LLAMAS demonstrate state-of-the-art performance and practical advantages, highlighting the method’s potential to enhance safe navigation in autonomous driving; future work includes more on-road data and dual-framework integration.

Abstract

Inspired by human driving focus, this research pioneers networks augmented with Focusing Sampling, Partial Field of View Evaluation, Enhanced FPN architecture and Directional IoU Loss - targeted innovations addressing obstacles to precise lane detection for autonomous driving. Experiments demonstrate our Focusing Sampling strategy, emphasizing vital distant details unlike uniform approaches, significantly boosts both benchmark and practical curved/distant lane recognition accuracy essential for safety. While FENetV1 achieves state-of-the-art conventional metric performance via enhancements isolating perspective-aware contexts mimicking driver vision, FENetV2 proves most reliable on the proposed Partial Field analysis. Hence we specifically recommend V2 for practical lane navigation despite fractional degradation on standard entire-image measures. Future directions include collecting on-road data and integrating complementary dual frameworks to further breakthroughs guided by human perception principles. The Code is available at https://github.com/HanyangZhong/FENet.

Figures (8)

• Figure 1: Skilled drivers focus their gaze far ahead on the road. The upper left image shows the full camera view, while the lower left image visualizes where experienced drivers look - far ahead along the road and lane lines.
• Figure 2: Architecture of the FENet lane detection framework, which uses a DLA34 backbone and Enhanced FPN. The input layers feed into internal layers integrated with positional non-local blocks (for FENetV1 or standard non-local blocks (for FENetV2) to capture spatial context. The internal layers connect to output layers that pass through Focusing Sampling and either standard IOU loss (for FENetV1) or Directional IOU loss (for FENetV2). FENetV1 (yellow pathway) and FENetV2 (blue pathway) share the common green pathway, with distinct supplementary modules.
• Figure 3: Visual depiction comparing Focusing Sampling (red dots) versus uniform sampling (blue dots). Focusing Sampling strategically emphasizes critical distant vanishing points along the lane while retaining informative nearby points. This accounts for perspective geometry, unlike standard uniform sampling that weights all regions equally.
• Figure 4: D-IoU Loss. D-IoU loss assesses lane prediction accuracy by comparing predicted lanes to ground truth, using the IoU of extended lane segments from sampled points along the lane. This directional, localized loss metric accurately reflects predict precision across the entire lane geometry.
• Figure 5: The proposed Partial Field of View metric enhances evaluation by subdividing the lower image half into distal fraction views after preprocessing cropping. Assessing model precision on the (a) top 1/2 field and (b) top 1/3 field aligns with driving gaze ahead needs. This practical metric gauges lane detection aptitude beyond existing methods.