基于轻量级MobileNet-SSD模型的人流量检测
HUMAN TRAFFIC DETECTION BASED ON LIGHTWEIGHT MOBILENET-SSD MODEL
张智 1盛健1
作者信息
- 1. 武汉科技大学计算机科学与技术学院 湖北武汉 430065;智能信息处理与实时工业系统湖北省重点实验室 湖北武汉 430065;武汉科技大学大数据科学与工程研究院 湖北武汉 430065
- 折叠
摘要
利用深度神经网络模型识别行人目标并检测具有十分高的价值.现实高密度行人检测场景中由于硬件基础以及网络性能消耗等问题的影响,常常需要选取处理速度高、硬件条件要求低的网络且同时兼顾视频监控的连续特性,因此选取了轻量级MobileNet-SSD网络来高效处理人头目标并引入帧间差分的方式,来有效对人头的椭圆特征目标进行跟踪;结合相关数理方法实现行人跨线计数的高性能人流量检测解决方案.在不同数据集上同现今一流的检测模型作比较,该方法皆表现出优良的检测性能.
Abstract
Using deep neural network models to identify and detect pedestrian targets has very high value.In the real high-density pedestrian detection scene,due to the impact of hardware foundation and network performance consumption,it is often necessary to select a network with high processing speed and low hardware requirements,while taking into account the continuous characteristics of video surveillance.Therefore,this paper selected the lightweight MobileNet-SSD network to efficiently process human head targets and introduced the method of inter-frame difference to effectively track the elliptical feature targets of the human head.The related mathematical methods were combined to achieve a high-performance pedestrian flow detection solution that counted pedestrians across the line.After comparing with the current first-class detection models on different data sets,the proposed method showed excellent detection performance.
关键词
MobileNet-SSD/帧间差分/跨线计数/轻量级网络Key words
MobileNet-SSD/Inter-frame difference/Cross-line count/Lightweight network引用本文复制引用
基金项目
国家自然科学基金项目(61673304)
国家社会科学基金重大计划项目(11&ZD189)
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据