all research
arXiv · cs.CV · Preprint · Under review

ENet-21: An Optimized Light CNN Structure for Lane Detection

Seyed Rasoul Hosseini, Hamid Taheri, Mohammad Teshnehlab Year 2024 arXiv 2403.19782
Read on arXiv arXiv:2403.19782DOI

Abstract

This study develops an optimized, lightweight CNN for lane detection that combines binary segmentation with Affinity Fields. Lane instances are recovered by clustering the segmentation and affinity outputs, which lets the method handle a varying number of lanes and lane-change scenarios without a fixed lane-count assumption. Using a less complex network than existing approaches, the method is demonstrated on the TuSimple dataset.

A study in efficient perception: getting robust lane understanding from a network small enough to be practical, without giving up flexibility on the number of lanes.

Approach

  • Binary segmentation + Affinity Fields. Pixels are segmented as lane / non-lane, while affinity fields encode which pixels belong to the same lane instance.
  • Clustering for instances. Lanes are recovered by clustering the segmentation and affinity outputs — so the method is not locked to a fixed lane count and naturally handles lane-change scenarios.
  • Lightweight by design. A less complex CNN than comparable approaches, evaluated on the TuSimple benchmark.