基于特征选取的监控运动目标识别模型研究

Research on surveillance motion target recognition model based on feature selection

根据深度学习网络模型,在YOLOv3框架基础上引进金字塔结构,构建基于金字塔特征选取的YOLOv3目标识别模型。利用深度残差网络提高精度,解决单纯模型加深过程中的退化问题,利用多尺度融合卷积机制实现大范围内小物体的全面检测,规避漏检情况。模型建立完成后进行综合实验,结果证明提出的目标识别模型与传统识别方法相比,在精度和准确性上都有明显提升,对于监控视频中运动目标的识别具有良好效果。

According to the deep learning network model, the pyramid structure is introduced on the basis of the YOLOv3 framework to build a YOLOv3 target recognition model based on pyramid feature selection. The deep residual network is used to improve the accuracy, solve the degradation problem in the process of deepening the simple model, and use the multi-scale fusion convolution mechanism to achieve comprehensive detection of small objects in a large range and circumvent the situation of missing detection. Comprehensive experiments are conducted after the model is established, and the results prove that the proposed target recognition model has significantly improved in accuracy and precision compared with traditional recognition methods, and has good effects on the recognition of moving targets in surveillance videos.