基於混合擴張卷積及卷積注意力模組的即時車道線偵測深度學習網路

Abstract

準確的車道偵測對於自動駕駛系統的安全運作至關重要。雖然 LaneATT（即時注意力引導車道偵測）等模型已經表現出強大的性能，但仍有改進其多尺度特徵擷取和優先處理關鍵車道資訊的能力。論文提出了透過整合混合擴張卷積（Hybrid Dilated Convolution, HDC）和卷積塊注意模組（Convolutional Block Attention Module, CBAM）對 LaneATT 模型進行改進。 HDC 模組以最小的運算成本實現多尺度特徵提取，而 CBAM 透過強調重要的空間和通道資訊來增強特徵圖。在 TuSimple 和 CULane 資料集上進行的大量實驗凸顯了我們方法的有效性，與原始 LaneATT 模型相比取得了卓越的性能。此外，消融實驗證實了 HDC 和 CBAM 能夠有效地擷取多尺度情境資訊並專注於相關特徵。Accurate lane detection is essential for the safe operation of autonomous driving systems. While models like LaneATT (Real-time Attention-guided Lane Detection ) have demonstrated strong performance, there is still room for improvement in their ability to extract multi-scale features and prioritize critical lane information. This paper proposes improvements to the LaneATT model by integrating Hybrid Dilated Convolution (HDC) and the Convolutional Block Attention Module (CBAM). The HDC module facilitates multi-scale feature extraction with minimal computational cost, while CBAM enhances feature maps by emphasizing important spatial and channel-wise information. Extensive experiments on the TuSimple and CULane datasets highlight the effectiveness of our approach, achieving superior performance compared to the original LaneATT model. Additionally, ablation experiment confirm that HDC and CBAM are able to effectively capture multi-scale contextual information and focus on relevant features.