孫傳龍 趙紅 崔翔宇 牟亮 徐福良 路來偉



摘要: 為提高無人駕駛汽車感知系統硬件資源利用率,構建了一種基于特征融合的無人駕駛多任務感知算法。采用改進的CSPDarknet53作為模型的主干網絡,通過構建特征融合網絡與特征融合模塊對多尺度特征進行提取與融合,并以7種常見道路物體的檢測與可行駛區域的像素級分割兩任務為例,設計多任務模型DaSNet(Detection and Segmentation Net)進行訓練與測試。使用BDD100K數據集對YOLOv5s、Faster R-CNN以及U-Net模型進行訓練,并對mAP、Dice系數以及檢測速度等性能指標做出對比分析。研究結果表明:DaSNet多任務模型在道路物體檢測任務上,mAP值分別比YOLOv5s和Faster RCNN高出0.5%和4.2%,在RTX2080Ti GPU上達到121FPS的檢測速度;在占優先權與不占優先權的可行駛區域上分割的Dice值相較于U-Net網絡分別高出了4.4%與6.8%,有較明顯的提升。
關鍵詞: 無人駕駛;多任務;特征融合;道路物體檢測;可行駛區域分割
中圖分類號: TP391文獻標識碼: A
Multi-task Sensing Algorithm for Driverless Vehicle Based on Feature Fusion
SUN Chuanlong1,ZHAO Hong1,CUI Xiangyu2,MU Liang1,XU Fuliang1,LU Laiwei1
Abstract:In order to improve the utilization of hardware resources of driverless vehicle perception system, a multi-task driverless vehicle perception algorithm based on feature fusion is constructed. The improved CSPDarknet53 is used as the backbone network of the model, and multi-scale features are extracted and fused by constructing feature fusion network and feature fusion module. The detection of 7 common road objects and pixel-level segmentation of the driving area are taken as examples. Multi-task DaSNet (Detection and Segmentation Net) is designed for training and testing. In order to compare model performance, BDD100K data set is used to train YOLOv5s, Faster R-CNN and U-NET models, and comparative analysis is made on mAP, Dice coefficient and detection speed and other performance indicators. The results showed that DaSNet multi-task model′s mAP value is 0.5% and 4.2% higher than YOLOv5s and Faster RCNN, respectively, and the detection speed of 121FPS can be achieved on RTX2080Ti GPU. Compared with U-NET network, Dice value of segmentation in priority and non-priority drivable are 4.4% and 6.8% higher, showing an obvious improvement.
Key words: driverless vehicle; multi-task; fature fusion; road object dection;driveable area segmentation
0 引言
目前,無人駕駛已經成為眾多國家的發展戰略之一,其中,感知系統[1-3]作為無人駕駛汽車中不可或缺的部分之一,對環境的感知適應性、實時性直接影響了無人駕駛汽車的安全性和可靠性,而目標檢測、語義分割作為感知系統中的兩大任務,其效果將直接影響無人駕駛汽車決策系統的決策質量。
近年來,深度學習[4]在目標檢測、語義分割領域的應用取得了重大的突破,這很大程度上歸功于大型數據集、計算機強大的計算能力、復雜的網絡架構和優化算法的進展。在目標檢測領域,兩階段算法……