基於語意分割以及權重損失函數的車道線偵測

Abstract

車道線偵測一直是自動駕駛與電腦視覺領域中的重要任務，並且多年來已有大量的研究推動其成長與進步。近期多數新提出的研究大多採用座標點偵測或曲線偵測的方法。本研究提出了一種不同於現有常見模型的車道檢測方法。本研究利用了 Vision Transformer (ViT），透過基於語意分割的模型來解決現實車道檢測中常見的困難場景，例如彎曲的車道標記與有陰影的環境。為了解決正負樣本極度不平均的問題，本研究提出使用權重二元交叉熵損失函數。實驗結果顯示，在廣泛使用的 TuSimple 數據集上，本研究所提出的模型在加入權重二元交叉熵損失函數的調整後從原先的64.2%進步到了 96.45% 的準確率，與目前的一線模型並駕齊驅。本研究所提出的損失函數不但在 DeepLabV3+ 模型上也提升 18% 的準確率，同時在更為複雜的道路環境如陰影遮蔽、大車流量等，也展現出顯著的表現。 本研究的模型更能順利偵測分岔車道線和斷裂車道線等，是其餘基於點偵測和基於曲線偵測模型容易出現誤偵測的困難情境。

Lane detection has long been a critical task in the fields of autonomous driving and computer vision, with extensive research and development contributing to its progress over the years. With most of the recently proposed models adopting point-based and curve-based methods, the study presents an alternative method of lane detection that differs from other state-of-the-art models. Taking advantage of the powerful vision transformer (ViT), this research aims to utilize a segmentation-based model to address challenging scenarios often encountered in real-world lane detection, such as curved lane markings and environments with shadows. To solve the problem of extreme class imbalance, a custom weighted binary cross-entropy loss is proposed. Experimental results of the widely adopted TuSimple dataset showed that the proposed model improved from 64.2% accuracy to achieving a favorable accuracy of 96.45%, on par with many recent models, The loss function also provided 18% accuracy improvement on DeepLabV3+, while simultaneously exhibiting substantial performance in more complex scenarios such as crowded and shadowed lanes. The proposed model also has the ability to predict forked lanes and disconnected lanes, which other point-based and curve-based methods fail to predict.