基于轻量级神经网络的RGB-D人体目标检测  

RGB-D Human Target Detection Based on Lightweight Neural Network

作　　者：谭方[1] 冯晓毅[1] 马玉鹏 TAN Fang;FENG Xiaoyi;MA Yupeng(School of Electronics and Information,Northwestern Polytechnical University,Xi an 710072,China)

机构地区：[1]西北工业大学电子信息学院,西安710072

出　　处：《微处理机》2022年第1期34-38,共5页Microprocessors

摘　　要：针对现有基于神经网络的人体目标检测算法网络结构复杂,运算量大,不利于实际应用,以及传统方法检测精度较差的问题,提出一种新的轻量级检测算法,使用无锚框机制,并将MobileNetV3作为主干网络。该网络支持多个数据输入方式,可分别以RGB彩色图、深度图或RGB-D作为输入。通过在两个公开数据集和自采集数据集中的试验证明,新算法总体检测精度及运行效率均优于已有算法,获得较为理想的每秒峰值速度(FLOPS)。在英特尔i5-7200 CPU平台下,以RGB-D和Depth为输入的帧率分别可达32f/s和55f/s,以RGB为输入的表现优于同级别轻量级网络YOLOV3-Tiny。Aiming at the problems of the existing human target detection algorithm based on neural network, such as complex network structure, large amount of computation, which is not conducive to practical application, and poor detection accuracy of traditional methods, a new lightweight detection algorithm is proposed, which uses anchor-free mechanism and uses MobileNetV3 as the backbone network. The network supports multiple data input modes, and RGB color map, Depth map or RGB-D can be used as input respectively. Experiments on two public data sets and self-collected data sets show that the overall detection accuracy and running efficiency of the new algorithm are better than the existing algorithms, and the ideal peak velocity per second(FLOPS) is obtained. On Intel i5-7200 CPU platform, the frame rate with RGB-D and Depth as input can reach 32 f/s and 55 f/s respectively, and the performance with RGB as input is better than that of YOLOV3-Tiny, a lightweight network of the same level.

关 键 词：RGB-D技术 行人检测 轻量级网络 神经网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]