Vehicle and lane detection algorithm based on MSFA-Net

doi:10.3969/j.issn.1674-8484.2024.03.017

Abstract

Abstract:

Vehicle detection and lane segmentation are important components of automatic driving sensing system, and their basic requirements are high precision and real-time. Therefore, a dual-task multi-scale feature aggregation network (MSFA-Net) was proposed, which was composed of one feature extraction network and two detection branch networks, and realized the simultaneous detection of vehicles and lane lines. First, E-ELAN network was used to construct the shared backbone feature network. Convolutional basic structure plus (CBS+) was designed for bottom-up feature fusion to improve accuracy in vehicle detection branch. To enhance the accuracy of discontinuous and nonlinear lane segmentation in lane segmentation branch, FeatFuse module was proposed for adaptive weight fusion of multi-features and context dilated convolutional basic structure (CDBS) for sampling fusion features through multi-dilation convolution of trapezoidal structure. The results show that on the BDD100K dataset, the average accuracy, recall rate and pixel accuracy of MSFA-Net reach 81.3%, 90.1% and 80.1% respectively, and the detection frame rate reaches 41.6 frames /s, which can better adapt to the needs of real-life driving scenarios.

Key words: vehicle detection, traffic image, deep learning, lane segmentation, dual task multi-scale feature aggregation network (MSFA-Net)

CLC Number:

TP391.4

WEN Bin, DING Yifu, HU Yiming, PENG Shun, HU Hui. Vehicle and lane detection algorithm based on MSFA-Net[J]. Journal of Automotive Safety and Energy, 2024, 15(3): 433-442.

Figures/Tables 19

References 21

[1]	CHEN Duanyu, LIN Yuhao, PENG Yangjie. Nighttime brake-light detection by nakagami imaging[J]. IEEE Trans Intel Transport Syst, 2012, 13(4): 1627-1637.
[2]	Srinivasa N. Vision-based vehicle detection and tracking method for forward collision warning in automobiles[C]// Intel Vehic Symp, 2002. IEEE, Versailles, France, 2002
[3]	HOU Changzheng, HOU Jin, YU Chaochao. An efficient lane markings detection and tracking method based on vanishing point constraints[C]// 2016 35th Chin Contr Conf (CCC), Chengdu, China, 2016.
[4]	Somawirata I K, Utaminingrum F. Road detection based on the color space and cluster connecting[C]// 2016 IEEE Int'l Conf Sign Imag Proc (ICSIP), Beijing, China. 2016.
[5]	金智林, 何麟煊, 赵万忠. 用于智能汽车的复杂光照环境车道线检测及跟踪方法[J]. 汽车安全与节能学报, 2019, 10(4): 459-466.
	JIN Zhilin, HE Linxuan, ZHAO Wanzhong. Detection and tracking method of lane line for intelligent vehicles under complex illumination condition[J]. J Autom Safe Energ, 2019, 10(4): 459-466. (in Chinese)
[6]	REN Shaoqing, HE Kaiming, Girshick R, et al. Faster R-CNN: Towards real-time object detection with region proposal networks[J]. IEEE Trans Patt Analy Mach Intel, 2017, 39(6): 1137-1149.
[7]	Bochkovskiy A, WANG Chienyao, LIAO Hongyuan M. YOLOv4: Optimal speed and accuracy of object detection[J]. ArXiv, 2020, abs/2004.10934.
[8]	GE Zheng, LIU Songtao, WANG Feng, et al. YOLOX: Exceeding YOLO Series in 2021[J]. ArXiv, 2021, abs/2107.08430.
[9]	WANG Chienyao, Bochkovskiy A, LIAO Hongyuan M. YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors[J]. ArXiv, 2022, abs/2207.02696.
[10]	丁飞, 米冠宇, 童恩, 等. 多通路高分辨率网络与注意力机制融合的车辆检测模型[J]. 汽车安全与节能学报, 2022, 13(1): 122-130.
	DING Fei, MI Guanyu, TONG En, et al. Multi-channel high-resolution network and attention mechanism fusion for vehicle detection model[J]. J Autom Safe Energ, 2022, 13(1): 122-130. (in Chinese)
[11]	Neven D, Brabandere B D, Georgoulis S, et al. Towards end-to-end lane detection: an instance segmentation approach[C]// 2018 IEEE Intel Vehic Symp (IV), Changshu, China. 2018
[12]	Lee S, Kim J, Yoon J S, et al. VPGNet: Vanishing point guided network for lane and road marking detection and recognition[C]// 2017 IEEE Int'l Conf Compu Visio (ICCV), Venice. 2017.
[13]	Teichmann M, Weber M, Zöllner J M, et al. MultiNet: Real-time joint semantic reasoning for autonomous driving[C]// 2018 IEEE Int't Vehic Symp (IV), Changshu, China. 2018.
[14]	QIAN Yeqiang, Dolan J M, YANG Ming. DLT-Net: Joint detection of drivable areas, lane lines, and traffic objects[J]. IEEE Trans Intel Transport Syst, 2020, 21(11): 4670-4679.
[15]	WU Dong, LIAO Manwen, ZHANG Weitian, et al. YOLOP: You only look once for panoptic driving perception[J]. Mach Intel Res, 2022, 19(6): 550-562.
[16]	Vu D, Ngo B, Phan H N. HybridNets: End-to-end perception network[J]. ArXiv, 2022, abs/2203.09035.
[17]	ZHAO Hengshuang, SHI Jianping, QI Xiaojuan, et al. Pyramid scene parsing network[C]// 2017 IEEE Conf Comput Visio Patt Recog (CVPR), Honolulu, HI. 2017.
[18]	LIN Tsungyi, Goyal P, Girshick R, et al. Focal loss for dense object detection[J]. IEEE Trans Patt Analy Mach Intel, 2020, 42(2): 318-27.
[19]	Gevorgyan Z. SIoU Loss: More powerful learning for bounding box regression[J]. ArXiv, 2022, abs/2205.12740.
[20]	LI Xiaoya, SUN Xiaofei, MENG Yuxian, et al. Dice loss for data-imbalanced NLP tasks[J]. ArXiv, 2019, abs/1911.02855.
[21]	YU Fisher, CHEN Haofeng, WANG Xin, et al. BDD100K: A diverse driving dataset for heterogeneous multitask learning[C] // 2020 IEEE/CVF Conf Comput Visio Patt Recog (CVPR), Seattle, WA, USA. 2020.

实验序号	YOLOP	主干网络	损失函数	CBS+	FeatFuse CDBS	Mosaic Mixup	mAP@0.5%	Rcall / %	PA / %	IoU / %	FPS帧/ s
1	√	—	—	—	—	—	76.5	89.2	70.5	26.2	37.0
2	√	√	—	—	—	—	78.4	89.6	70.6	26.7	—
3	√	√	√	—	—	—	80.2	89.5	75.9	25.2	—
4	√	√	√	√	—	—	80.3	89.6	77.5	25.7	—
5	√	√	√	√	√	—	81.0	89.3	79.0	27.2	—
6	√	√	√	√	√	√	81.3	90.1	80.1	26.0	41.6

实验序号	YOLOP	主干网络	损失函数	CBS+	FeatFuse CDBS	Mosaic Mixup	mAP@0.5%	Rcall / %	PA / %	IoU / %	FPS帧/ s
1	√	—	—	—	—	—	76.5	89.2	70.5	26.2	37.0
2	√	√	—	—	—	—	78.4	89.6	70.6	26.7	—
3	√	√	√	—	—	—	80.2	89.5	75.9	25.2	—
4	√	√	√	√	—	—	80.3	89.6	77.5	25.7	—
5	√	√	√	√	√	—	81.0	89.3	79.0	27.2	—
6	√	√	√	√	√	√	81.3	90.1	80.1	26.0	41.6

模型	mAP@0.5 / %	Recall / %
MultiNet	60.2	81.3
DLT-Net	68.4	89.4
YOLOV5s	77.2	86.8
YOLOP	76.5	89.2
YOLOX	77.9	82.3
YOLOV7	80.1	90.2
MSFA-Net	81.3	90.1

模型	mAP@0.5 / %	Recall / %
MultiNet	60.2	81.3
DLT-Net	68.4	89.4
YOLOV5s	77.2	86.8
YOLOP	76.5	89.2
YOLOX	77.9	82.3
YOLOV7	80.1	90.2
MSFA-Net	81.3	90.1

模型	PA / %	IoU / %
ENet	34.1	14.6
SCNN	35.8	15.8
ENet-SAD	36.6	16.0
PIDNet	55.8	21.8
FFNet	57.1	22.2
YOLOP	70.5	26.2
MSFA-Net	80.1	26.0