基于轻量化CenterNet的智能车辆目标检测算法

doi:10.12141/j.issn.1000-565X.230513

摘要/Abstract

摘要：

针对当前目标检测算法参数较多，计算量较大，导致响应速度较慢，难以推广应用于智能车辆系统的问题，提出一种改进的CenterNet目标检测算法。即应用轻量化MobileNetV3网络替换原ResNet-50网络，降低计算量；应用深度可分离的PANet替换特征增强网络，获得多尺度特征信息融合后的特征，并引入SimAM注意力机制在特征融合前强化目标特征关注度，再用SiLU激活函数替换原目标检测网络中的ReLU激活函数，增强网络学习能力。最后提出CPAN-ASFF模块对深度可分离的PANet输出多尺度特征图进行融合，提高目标检测精度。应用优化后的KITTI数据集对改进后的CenterNet目标检测算法进行训练及检测验证，结果表明：其平均准确率为80.7%，比原始CenterNet目标检测算法提高了12个百分点，其检测速度为65 f/s，其参数量为8.91 M，较原算法减少72.73%，改进后的算法在遮挡目标、重叠目标以及与背景相似目标的检测效果上表现更优。且在SODA10M数据集中，文中提出的算法的检测精度与速度也都优于当前主流算法。该研究对算法的优化及实验为智能车辆在实际工程中的应用奠定了技术支撑。

关键词: 智能车辆, 目标检测, 无先验框, CenterNet, 轻量化, 深度可分离卷积

Abstract:

Aiming at the problem that the current object detection algorithm has many parameters and larger computation amount, resulting in slower response and difficulties of application in intelligent vehicle systems, this paper proposed an improved CenterNet object detection algorithm. That is, the lightweight MobileNetV3 network was applied to replace the original ResNet-50 network to reduce the amount of computation; the depth-separable PANet was applied to replace the feature enhancement network to obtain the features after the fusion of the multi-scale feature information, and the SimAM attention mechanism was introduced to strengthen the attention of the target features before the fusion of the features, then the SiLU activation function was used to replace the original object detection network in the ReLU activation function to enhance the network learning. Finally, the CPAN-ASFF module was proposed to fuse the depth-separable PANet output multi-scale feature maps to improve the object detection accuracy. The optimized KITTI dataset was applied to train and detect the improved CenterNet object detection algorithm for validation. And the results show that its mean average precision is 80.7%, which is 12 percentage points higher than that of the original CenterNet object detection algorithm; its detection speed is 65 f/s, and its number of parameters is 8.91 M, which is 72.73% less than the original algorithm. The improved algorithm performs better in the detection of occluded objects, overlapped objects and objects similar to the background. In the SODA10M dataset, the detection accuracy and speed of the algorithm are better than the current mainstream algorithms. The optimization of the algorithm and the experiments in the paper laid the technical support for the application of intelligent vehicles in practical engineering.

Key words: intelligent vehicles, object detection, anchor-free, CenterNet, lightweight, depthwise separable convolution

中图分类号:

TP391.4

岳永恒, 宁睿厚. 基于轻量化CenterNet的智能车辆目标检测算法[J]. 华南理工大学学报(自然科学版), 2024, 52(8): 45-55.

YUE Yongheng, NING Ruihou. Intelligent Vehicle Object Detection Algorithm Based on Lightweight CenterNet[J]. Journal of South China University of Technology(Natural Science Edition), 2024, 52(8): 45-55.

图/表 15

图1

图2

表1

图3

图4

图5

图6

图7

图8

图9

图10

图11

表2

消融试验"

方法	Mobile-netV3	FPN	CPAN-ASFF	SiLU	SimAM	$s A P, C / %$	$s A P, P / %$	参数/M	F/（f·s^-1）	$M A P / %$
1	×	×	×	×	×	78.9	58.5	32.67	68	68.7
2	√	×	×	×	×	78.6	53.6	4.24	82	66.1
3	√	√	×	×	×	84.2	59.6	5.16	76	71.9
4	√	√	√	×	×	89.8	67.6	8.91	66	78.7
5	√	√	√	√	×	90.0	69.0	8.91	66	79.5
6	√	√	√	√	√	90.6	70.9	8.91	65	80.7

表2

图12

表3

不同模型在数据集上检测结果"

网络模型	改进KITTI		SODA10M		参数/M
网络模型	$M A P / %$	$F / f · s - 1$	$M A P / %$	$F / f · s - 1$	参数/M
Faster-RCNN	74.3	30	55.7	28	137.10
YOLOv3	72.5	64	54.6	64	61.95
YOLOV4	76.1	52	57.1	50	64.36
YOLOV5x	77.6	44	59.3	43	87.78
CenterNet	68.7	68	53.8	67	32.67
文中提出的算法	80.7	65	63.3	64	8.91

表3

参考文献 33

1	GIRSHICK R， DONAHUE J， DARRELL T，et al ．Rich feature hierarchies for accurate object detection and semantic segmentation［C］∥ Proceedings of 2014 IEEE Conference on Computer Vision and Pattern Recognition（CVPR）．Columbus：IEEE，2014：580-587.
2	GIRSHICK R， DONAHUE J， DARRELL T，et al ．Region-based convolutional networks for accurate object detection and segmentation［J］．IEEE Transactions on Pattern Analysis and Machine Intelligence，2015，38（1）：142-158.
3	GIRSHICK R ．Fast R-CNN［C］∥ Proceedings of 2015 IEEE International Conference on Computer Vision （ICCV）．Santiago：IEEE，2015：1440-1448.
4	HE K M， ZHANG X Y， REN S Q，et al ．Spatial pyramid pooling in deep convolutional networks for visual re-cognition［J］．IEEE Transactions on Pattern Analysis and Machine Intelligence，2015，37（9）：1904-1916.
5	REN S Q， HE K M， GIRSHICK R，et al ．Faster R-CNN：towards real-time object detection with region proposal networks［J］．Advances in Neural Information Processing Systems，2017，39（6）：1137-1149.
6	REDMON J， DIVVALA S， GIRSHICK R，et al ．You only look once：unified，real-time object detection［C］∥ Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition （CVPR）．Las Vegas：IEEE，2016：779-788.
7	LIU W， ANGUELOV D， ERHAN D，et al ．Ssd：single shot multibox detector［C］∥ Proceedings of 2016 European Conference on Computer Vision （ECCV）．Amsterdam：Springer，2016：21-37.
8	SONG Y K， HONG S H， HU C T，et al ．MEB-YOLO：an efficient vehicle detection method in complex traffic road scenes［J］．Computers Materials & Continua，2023，75（3）：5761-5784.
9	CAO G M， XIE X M， YANG W Z，et al ．Feature-fused SSD：fast detection for small objects［C］∥ Proceedings of Ninth International Conference on Graphic and Image Processing （ICGIP 2017）．Qingdao：SPIE，2018：381-388.
10	LIN T Y， GOYAL P， GIRSHICK R，et al ．Focal loss for dense object detection［C］∥ Proceedings of 2017 IEEE International Conference on Computer Vision （ICCV）．Venice：IEEE，2017：2980-2988.
11	ZHAO G C， PANG J， ZHANG H，et al ．Anchor-free network for multi-class object detection in remote sensing images［C］∥ Proceedings of 2020 39th Chinese Control Conference （CCC）．Shenyang：IEEE，2020：7510-7515.
12	LAW H， DENG J ．CornerNet：detecting objects as paired keypoints［C］∥ Proceedings of 2018 European Conference on Computer Vision （ECCV）．Munich：Springer，2018：734-750.
13	CARION N， MASSA F， SYNNAEVE G，et al ．End-to-end object detection with transformers［C］∥ Proceedings of 2020 European Conference on Computer Vision （ECCV）．Glasgow：Springer，2020：213-229.
14	ZHOU X Y， WANG D Q， KRÄHENBÜHL P ．Objects as points［J］．arXiv preprint arXiv：，2019.
15	HOWARD A， SANDLER M， CHU G，et al ．Searching for mobilenetv3［C］∥ Proceedings of 2019 IEEE International Conference on Computer Vision （ICCV）．Seoul：IEEE，2019：1314-1324.
16	LIU S， QI L， QIN H F，et al ．Path aggregation network for instance segmentation［C］∥ Proceedings of 2018 IEEE Conference on Computer Vision and Pattern Recognition （CVPR）．Salt Lake City：IEEE，2018：8759-8768.
17	YANG L X， ZHANG R Y， LI L D，et al ．Simam：a simple，parameter-free attention module for convolutional neural networks［C］∥ Proceedings of International Conference on Machine Learning （ICML）．Williamstown：PMLR，2021：11863-11874.
18	RAMACHANDRAN P， ZOPH B， LE Q V ．Searching for activation functions［J］．arXiv preprint arXiv：，2017.
19	GLOROT X， BORDES A， BENGIO Y ．Deep sparse rectifier neural networks［C］∥ Proceedings of Fourteenth International Conference on Artificial Intelligence and Statistics．Fort Lauderdale：JMLR，2011：315-323.
20	GEIGER A， LENZ P， URTASUN R ．Are we ready for autonomous driving？the KITTI vision benchmark suite［C］∥ Proceedings of 2012 IEEE Conference on Computer Vision and Pattern Recognition （CVPR）．Providence：IEEE，2012：3354-3361.
21	NEUBECK A， VAN GOOL L ．Efficient non-maximum suppression［C］∥ Proceedings of 18th International Conference on Pattern Recognition （ICPR）．Hong Kong：IEEE，2006：850-855.
22	LIN T Y， DOLLÁR P， GIRSHICK R，et al ．Feature pyramid networks for object detection［C］∥ Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition （CVPR）．Honolulu：IEEE，2017：2117-2125.
23	ZHANG X X， LI N， ZHANG R X ．An improved lightweight network MobileNetv3 based YOLOv3 for pedestrian detection［C］∥ Proceedings of 2021 IEEE International Conference on Consumer Electronics and Computer Engineering （ICCECE）．Guangzhou：IEEE，2021：114-118.
24	WANG S Y， ZHAO J，TA N，et al ．A real-time deep learning forest fire monitoring algorithm based on an improved Pruned+ KD model［J］．Journal of Real-Time Image Processing，2021，18（6）：2319-2329.
25	HOWARD G， ZHU M L， CHEN B，et al ．Mobilenets：efficient convolutional neural networks for mobile vision applications［J］．arXiv preprint arXiv：，2017.
26	SANDLER M， HOWARD A， ZHU M L，et al ．Mobilenetv2：inverted residuals and linear bottlenecks［C］∥ Proceedings of 2018 IEEE Conference on Computer Vision and Pattern Recognition （CVPR）．Salt Lake City：IEEE，2018：4510-4520.
27	HU J， SHEN L， SUN G ．Squeeze-and-excitation networks［C］∥ Proceedings of 2018 IEEE Conference on Computer Vision and Pattern Recognition （CVPR）．Salt Lake City：IEEE，2018：7132-7141.
28	ELFWING S， UCHIBE E， DOYA K ．Sigmoid-weighted linear units for neural network function approximation in reinforcement learning［J］．Neural Networks，2018，107：3-11.
29	MA N N， ZHANG X Y， LIU M，et al ．Activate or not：learning customized activation［C］∥ Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）．Nashville：IEEE，2021：8032-8042.
30	LIU S T， HUANG D， WANG Y H ．Learning spatial fusion for single-shot object detection［J］．arXiv preprint arXiv：，2019.
31	GUO M H， XU T X， LIU J J，et al ．Attention mechanisms in computer vision：a survey［J］．Computational Visual Media，2022，8（3）：331-368.
32	GEIGER A， LENZ P， STILLER C，et al ．Vision meets robotics：the KITTI dataset［J］．The International Journal of Robotics Research，2013，32（11）：1231-1237.
33	OKSUZ K， CAM B C， KALKAN S，et al ．Imbalance problems in object detection：a review［J］．IEEE Transactions on Pattern Analysis and Machine Intelligence，2020，43（10）：3388-3415.

Input	Module/numbers	K	SE	Activation
512×512×3	Conv2D/1	3	×	HS
256×256×16	Bneck/1	3	×	RE
256×256×16	Bneck/2	3	×	RE
128×128×24	Bneck/3	5	√	RE
64×64×40	Bneck/4	3	×	HS
32×32×80	Bneck/2	3	√	HS
32×32×112	Bneck/3	5	√	HS

[1]	杜启亮, 向照夷, 田联房. 面向嵌入式设备的扶梯客流量实时统计方法[J]. 华南理工大学学报(自然科学版), 2022, 50(6): 60-70.
[2]	符锌砂, 曾彦杰, 马丽, 等. 基于轻量化神经网络的公路监控场景天气识别研究[J]. 华南理工大学学报(自然科学版), 2022, 50(3): 1-8.
[3]	易清明, 吕人毅, 石敏, 等. 融合多尺度空洞卷积与反卷积的轻量化目标检测[J]. 华南理工大学学报(自然科学版), 2022, 50(12): 41-48.
[4]	张香竹, 张立家, 宋逸凡, 等. 基于深度学习的无人机单目视觉避障算法[J]. 华南理工大学学报(自然科学版), 2022, 50(1): 101-108, 131.
[5]	谭光李昌镐詹昭焕. 基于端云协同的目标检测跟踪自适应调度算法[J]. 华南理工大学学报（自然科学版）, 2021, 49(7): 86-93.
[6]	李波饶浩波. 复杂场景下特征增强的显著性目标检测方法[J]. 华南理工大学学报（自然科学版）, 2021, 49(11): 135-144.
[7]	吴秋霞黎玲敏. 基于点云俯视图重映射的 3D 目标检测[J]. 华南理工大学学报(自然科学版), 2021, 49(1): 39-46.
[8]	蔡力亚, 赵克刚, 李剑峰, 等. 短玻纤增强复合材料的性能及其在车身上的应用[J]. 华南理工大学学报（自然科学版）, 2020, 48(3): 108-115,135.
[9]	朱斌田联房杜启亮余陆斌. 基于自适应双分数阶光流模型的运动目标分割 [J]. 华南理工大学学报(自然科学版), 2019, 47(2): 77-84.
[10]	韩雷刚范文艳王国河王振民. 水下焊接机器人系统轻量化全密封送丝装置的研制 [J]. 华南理工大学学报（自然科学版）, 2018, 46(5): 135-140.
[11]	刘玉杰曹先知李宗民李华. 结合关联特征和卷积神经网络的视频目标检测[J]. 华南理工大学学报（自然科学版）, 2018, 46(12): 26-33.
[12]	左文杰马洪刚桂春阳李行. 动静刚度约束下客车骨架的轻量化设计[J]. 华南理工大学学报(自然科学版), 2017, 45(8): 7-12.
[13]	符锌砂王祥波李海峰孟庆昕. 基于背景差分的高速公路运动目标检测算法[J]. 华南理工大学学报（自然科学版）, 2015, 43(4): 1-6.
[14]	许伦辉胡三根罗强周勇. 基于驾驶员类型的车辆换道模型[J]. 华南理工大学学报（自然科学版）, 2014, 42(8): 104-111.
[15]	王世明徐建闽罗强李日涵. 面向高速公路的车辆换道安全预警模型研究[J]. 华南理工大学学报（自然科学版）, 2014, 42(12): 40-50.