基于改进神经网络的视频序列运动目标识别方法  

Method of video sequence motion target recognition based on improved neural network

作　　者：范建伟[1] 李琳[1] 靳志鑫[1] FAN Jianwei;LI Lin;JIN Zhixin(Taiyuan University of Science and Technology,Taiyuan 030024,China)

机构地区：[1]太原科技大学,山西太原030024

出　　处：《现代电子技术》2024年第20期118-122,共5页Modern Electronics Technique

摘　　要：为改善运动目标检测效果,降低目标漏检率,提出一种基于改进神经网络的视频序列运动目标识别方法。构建改进YOLOv3的运动目标识别模型,以不同帧视频图像为模型输入,经过卷积层的初步特征提取后,输入到由5个残差模块组成的深层网络中。通过以上采样方式构建特征金字塔,实现对运动目标四尺度特征的捕捉。在特征金字塔的每一层,应用K-means算法对运动目标真实框进行聚类,确保候选框尺寸和比例与真实运动目标的统计特性相匹配;再利用获得的候选框和分类器对特征图上每个位置进行目标检测,运用非极大值抑制技术剔除重叠框,将斥力损失函数引入到网络训练总损失之中,使预测框无限贴近运动目标真实框,实现对运动目标的精准识别。实验结果表明,所提方法具有显著的运动目标识别能力,当聚类数目为12时,运动目标识别的AUC、F1指标可达到0.92、0.90,且计算量较少。A method of video sequence motion target recognition based on improved neural network is proposed to improve the motion target detection and reduce target missed-detection rate.An improved YOLOv3 motion target recognition model is constructed,which takes different frame video images as model input.After preliminary feature extraction by convolutional layers,the model is input into the deep network composed of 5 residual modules.The feature pyramid is constructed by means of the above sampling method to capture the four scale features of the motion target.At each layer of the feature pyramid,K-means algorithm is used to cluster the real boxes of the motion target,ensuring that the size and proportion of the candidate boxes match the statistical characteristics of the real motion target.The obtained candidate boxes and classifiers are used to detect targets at each position on the feature map.The non-maximum suppression technology is used to remove overlapping boxes,and the repulsive loss function is introduced into the total loss of network training to make the predicted boxes infinitely close to the real boxes of the motion target,so as to realize the precise recognition of sports goals.The experimental results show that the proposed method has significant ability in motion target recognition,when the number of clusters is 12,the AUC and F1 score indicators for motion target recognition can reach 0.92 and 0.90,and the computation is less.

关 键 词：视频序列 运动目标识别 改进YOLOv3网络 特征金字塔 K-MEANS算法 候选框聚类 

分 类 号：TN911.23-34[电子电信—通信与信息系统] TP391.41[电子电信—信息与通信工程]