车前运动目标检测是计算机视觉领域重要的研究方向，也是辅助驾驶系统重要的组成部分。由于辅助驾驶场景复杂，目标存在运动遮挡及运动模糊问题，影响目标检测精度和速度。因此，提高车前运动目标检测算法速度和精度具有重要意义。

针对车载高清视频冗余引起的计算资源浪费和检测速度慢的问题，提出一种光流网络与改进YOLOv3融合的车前运动目标检测算法。利用通道分割优化GhostNet网络，将分割后的特征图分别输入Bottleneck模块的两个分支，对分支1采用Ghost Moduld和深度可分离卷积提取深层特征，将分支1的深层特征图和分支2的特征图进行通道混洗，采用激活函数Leake Relu优化GhostNet网络，将优化后的GhostNet代替YOLOv3算法原始的骨干网络Darknet-53。对骨干网络输出的三个尺度特征分别引入三个自适应权重参数，进行多尺度自适应融合。对非关键帧使用光流网络LiteFlowNet预测与关键帧的光流场，关键帧特征通过光流场映射得到非关键帧特征，将关键帧特征和非关键帧特征输入分类定位网络，得到车前运动目标检测结果。

基于KITTY和BDD100k数据集对改进算法进行验证。相比原始YOLOv3算法，改进后的YOLOv3算法模型fps提高5.32倍，mAP达到54.16%。融合后算法通过对60s时长，帧率为33fps的视频进行检测，检测速度相对原始YOLOv3提高8.64倍，达到121fps，60s视频检测时长为16.36s，平均精度达到57.51%，模型最终大小为93M。测试结果表明，光流网络与改进的YOLOv3融合算法在保持模型精度的前提下，有效提高运动目标检测速度。

Motion Detection in front of vehicle is an important research direction in the field of computer vision, and it is also an important part of assistant driving system. Due to the complexity of assisted driving scene, there are motion occlusion and motion blur problems, which affect the accuracy and speed of target detection. Therefore, it is of great significance to improve the speed and accuracy of motion detection algorithm.

Aiming at the problem of waste of computing resources and slow detection speed caused by redundancy of vehicle HD video, motion detection algorithm in front of vehicle based on optical flow network and improved yolv3 is proposed. Channel split is used to optimize the GhostNet network. The segmented feature maps are input into the two branches of the pottleneck module. The deep features of branch 1 are extracted by ghost module and deep separable convolution. The deep feature maps of branch 1 and branch 2 are channel shuffled. The activation function Leake relu is used to optimize the GhostNet network, The optimized GhostNet is used to replace the original backbone network darknet-53 of yoov3 algorithm. Three adaptive weight parameters are introduced to the three scale features of the backbone network output for multi-scale adaptive fusion. For non key frames, the optical flow field of key frames is predicted by optical flow network liteflownet. Key frame features are mapped by optical flow field to obtain non key frame features. Key frame features and non key frame features are input into classification and positioning network to obtain motion detection results in front of vehicle.

The improved algorithm is verified based on KITTY and BDD100k datasets. Compared with the original algorithm, the fps of the improved algorithm is increased by 5.32 times, and the mAP is 54.16%. After fusion, the detection speed of the algorithm is 8.64 times higher than that of the original yolov3, reaching 121fps. The detection time of 60s video is 16.36s, and the mAP is 57.51%. The final size of the model weight is 93m. The test results show that the fusion algorithm of optical flow network and improved yolov3 can effectively improve the speed of motion detection while maintaining the accuracy of the model.

[1]Choi S, Lee K, Kim K, et al. Lane departure warning system using deep learning[J]. Journal of the Korea Industrial Information Systems Research, 2019, 24(2): 25-31.

[2]Zhu M, Wang X, Hu J. Impact on car following behavior of a forward collision warning system with headway monitoring[J].Transportation research part C: emerging technologies, 2020, 111: 226-244.

[3]Jang C, Kim C, Lee S, et al. Re-plannable automated parking system with a standalone around view monitor for narrow parking lots[J]. IEEE Transactions on Intelligent Transportation Systems, 2019, 21(2): 777-790.

[4]Lee S H, Lee S, Kim M H. Development of a driving behavior-based collision warning system using a neural network[J]. International journal of automotive technology, 2018, 19(5): 837-844.

[5]Sourial K, Kearns T, Carman T. COVID-19 Pandemic: Can We Adapt a System-wide Warfarin Monitoring Program and Maintain Safety[J]. Journal of Vascular Surgery: Venous and Lymphatic Disorders, 2021, 9(2): 538-539.

[6]Li D, Zhao D, Zhang Q, et al. Reinforcement learning and deep learning based lateral control for autonomous driving[J]. IEEE Computational Intelligence Magazine, 2019, 14(2): 83-98.

[7]Li D, Zhao D, Zhang Q, et al. Reinforcement learning and deep learning based lateral control for autonomous driving [application notes][J]. IEEE Computational Intelligence Magazine, 2019, 14(2): 83-98.

[8]Yahya S, Moghavvemi M, Almurib H A F, et al. A LabVIEW module to promote undergraduate research in control of AC servo motors of robotics manipulator[J]. Computer Applications in Engineering Education, 2020, 28(1): 139-153.

[9]江泽涛,覃露露.一种基于U-Net生成对抗网络的低照度图像增强方法[J].电子学报,2020,48(2):258-264.

[10]方路平,何杭江,周国民.目标检测算法研究综述[J].计算机工程与应用,2018,54(13):18-33.

[11]闫贺,黄佳,李睿安,等.基于改进快速区域卷积神经网络的视频SAR运动目标检测算法研究[J].电子与信息学报,2021,43(3):615-622.

[12]张国平,高兆彬.基于改进混合高斯模型的运动目标检测方法[J].计算机系统应用,2017(4):97-99.

[13]Bai Z, Gao Q, Yu X. Moving object detection based on adaptive loci frame difference method[C]//2019 IEEE international conference on mechatronics and automation (ICMA). IEEE, 2019: 2218-2223.

[14]苏佳,高丽慧.采用帧差法和相关滤波改进的TLD跟踪算法[J].计算机工程与设计,2020,41(06):1694-1700.

[15]Haddad B M, Dodge S F, Karam L J, et al. Locally Adaptive Statistical Background Modeling With Deep Learning-Based False Positive Rejection for Defect Detection in Semiconductor Units[J]. IEEE Transactions on Semiconductor Manufacturing, 2020, 33(3): 357-372.

[16]朱文杰,王广龙,田杰,等.空时自适应混合高斯模型复杂背景运动目标检测[J].北京理工大学学报, 2018, 38(2): 165-172.

[17]Zhou A, Xie W, Pei J. Background modeling in the fourier domain for maritime infrared target detection[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2019, 30(8): 2634-2649.

[18]Tocker Y, Hagege R R, Francos J M. Dynamic Spatial Predicted Background[J]. IEEE Transactions on Image Processing, 2020, 29: 5517-5530.

[19]Almatrafi M, Hirakawa K. Davis camera optical flow[J]. IEEE Transactions on Computational Imaging, 2019, 6: 396-407.

[20]Bao G, Li D, Mei Y. Key Frames Extraction Based on Optical-flow and Mutual Information Entropy[C]//Journal of Physics: Conference Series. IOP Publishing, 2020, 1646(1): 012112.

[21]Mei L, Lai J, Xie X, et al. Illumination-invariance optical flow estimation using weighted regularization transform[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2019, 30(2): 495-508.

[22]Manbari Z, AkhlaghianTab F, Salavati C. Hybrid fast unsupervised feature selection for high-dimensional data[J]. Expert Systems with Applications, 2019, 124: 97-118.

[23]Zhou X, Yang C, Yu W. Moving object detection by detecting contiguous outliers in the low-rank representation[J]. IEEE transactions on pattern analysis and machine intelligence, 2012, 35(3): 597-610.

[24]Zhang Y, Zhu D, Wang P, et al. Vision-based vehicle detection for VideoSAR surveillance using low-rank plus sparse three-term decomposition[J]. IEEE Transactions on Vehicular Technology, 2020, 69(5): 4711-4726.

[25]Krizhevsky A, Sutskever I, Hinton G E. Imagenet classification with deep convolutional neural networks[J]. Advances in neural information processing systems, 2012, 25: 1097-1105.

[26]Girshick R, Donahue J, Darrell T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2014: 580-587.

[27]He K, Zhang X, Ren S, et al. Spatial pyramid pooling in deep convolutional networks for visual recognition[J]. IEEE transactions on pattern analysis and machine intelligence, 2015, 37(9): 1904-1916.

[28]Li J, Liang X, Shen S M, et al. Scale-aware fast R-CNN for pedestrian detection[J]. IEEE transactions on Multimedia, 2017, 20(4): 985-996.

[29]Ren S, He K, Girshick R, et al. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks[J]. IEEE transactions on pattern analysis and machine intelligence,2017, 39(6):1137-1149.

[30]Liu W, Anguelov D, Erhan D, et al. SSD: Single shot multibox detector[C]//European conference on computer vision. Springer, Cham, 2016: 21-37.

[31]Lu S, Wang B, Wang H, et al. A real-time object detection algorithm for video[J]. Computers & Electrical Engineering, 2019, 77: 398-408.

[32]丁丹丹,吴熙林,佟骏超,等.联合时域虚拟帧的多帧视频质量增强方法[J].计算机辅助设计与图形学学报,2020,32(05):780-786.

[33]Han W, Khorrami P, Paine T L, et al. Seq-nms for video object detection[J]. arXiv preprint arXiv:1602.08465, 2016.

[34]Kang K, Li H, Yan J, et al. T-cnn: Tubelets with convolutional neural networks for object detection from videos[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2017, 28(10): 2896-2907.

[35]康金忠,王桂周,何国金,等.遥感视频卫星运动车辆目标快速检测[J].遥感学报,2020,24(09):1099-1107.

[36]唐悦,吴戈.基于灰度投影运动估计的ViBe改进算法[J].长春理工大学学报(自然科学版),2021,44(01):95-101.

[37]马庆禄,唐小垚.融合时域和分水岭信息的车辆检测算法[J].计算机工程与应用,2021,04(01):1-10.

[38]梁栋,何佳,石陆魁,等.结合运动目标检测和ResNet的车速车型智能识别[J].北京交通大学学报,2019,43(05):10-19.

[39]Gong L, Wang C. Research on Moving Target Tracking Based on FDRIG Optical Flow[J]. Symmetry, 2019, 11(9): 11-22.

[40]Yu X, Chen X, Huang Y, et al. Radar moving target detection in clutter background via adaptive dual-threshold sparse Fourier transform[J]. IEEE Access, 2019, 7: 58200-58211.

[41]柳长源,陈兰萍,Ersoy.基于车前灯的夜间车辆视频检测研究[J].计算机工程与应用,2019,55(02):253-258.

[42]乔瑞萍,王方,董员臣,等.基于改进HOG-LBP特征车前多目标分类仿真[J].计算机仿真,2020,37(11):138-141.

[43]祁海军,刘智嘉,赵金博,等.基于红外智能视频检测的自适应加速检测算法[J].激光与红外,2020,50(10):1212-1217.

[44]Coulter R, Han Q L, Pan L, et al. Data-driven cyber security in perspective—Intelligent traffic analysis[J]. IEEE transactions on cybernetics, 2019, 50(7): 3081-3093.

[45]郑远攀,李广阳,李晔.深度学习在图像识别中的应用研究综述[J].计算机工程与应用, 2019, 55(12): 20-36.

[46]Lin W, He X, Dai W, et al. Key-point sequence lossless compression for intelligent video analysis[J]. IEEE MultiMedia, 2020, 27(3): 12-22.

[47]张有健,陈晨,王再见.深度学习算法的激活函数研究[J].无线电通信技术,2021,47(01):115-120.

[48]Wang P, Zheng C, Xiong S. Hyperspectral Image Dimensionality Reduction via Graph Embedding in Core Tensor Space[J]. IEEE Geoscience and Remote Sensing Letters, 2020,99:1-5.

[49]Wang S, Tang W, Liu C, et al. Digital Image Correlation Measurement of the Deformation and Failure in PBX Brazilian Discs Reinforced with CFRP Patches[J]. Propellants, Explosives, Pyrotechnics, 2021, 46(4): 548-554.

[50]Tanaka M. Weighted sigmoid gate unit for an activation function of deep neural network[J]. Pattern Recognition Letters, 2020, 135: 354-359.

[51]Yang X, Li X, Guan Y, et al. Overfitting reduction of pose estimation for deep learning visual odometry[J]. China Communications, 2020, 17(6): 196-210.

[52]郭玥秀,杨伟,刘琦,等.残差网络研究综述[J].计算机应用研究,2020,37(05):1292-1297.

[53]肖振久,杨晓迪,魏宪,等.改进的轻量型网络在图像识别上的应用[J].计算机科学与探索,2021,15(04):743-753.

[54]Zhao X, Zhang Y, Zhang T, et al. Channel splitting network for single MR image super-resolution[J]. IEEE Transactions on Image Processing, 2019, 28(11): 5649-5662.

[55]Peng H, Zhang Y, Yang S, et al. Battlefield Image Situational Awareness Application Based on Deep Learning[J]. IEEE Intelligent Systems, 2019, 35(1): 36-43.

[56]周海赟,项学智,翟明亮,等.结合注意力机制的深度学习光流网络[J].计算机科学与探索,2020,14(11):1920-1929..