轻量级车辆行人检测模型研究及Android部署
Research on lightweight vehicle pedestrian detection model and Android deployment
王道斌 1李宸翔 1严运兵1
作者信息
- 1. 武汉科技大学汽车与交通工程学院,武汉 430065
- 折叠
摘要
针对目标检测模型参数量大,难以部署在移动端设备的问题,提出了一种轻量化车辆行人检测模型YOLOv8-TI(traffic information).设计全新的轻量级参数共享SPG Detect检测头以降低模型的参数量和计算量;提出全局平衡通道路径聚合网络(GBC-PAN)结构,平衡网络通道数量,通过跨尺度的加权链接,实现了自顶向下和自底向上的双向特征融合;此外,引入动态非单调聚焦机制的损失函数(Wise Loss)代替原损失函数以提升预测框精度.实验结果发现,提出的目标检测模型YOLOv8-TI在保持较高精度的同时,参数量、计算量和模型体积分别为YOLOv8n的52.1％、58.0％和54％.通过与其他轻量级目标检测模型对比,验证了该方法的有效性和卓越性.将YOLOv8-TI进行Android移动端部署,在荣耀20和荣耀80 GT上进行了测试,FPS可达24帧和31帧,满足实时性需求,有望进一步集成在自动驾驶汽车上完成交通信息检测功能.
Abstract
Object detection models usually have a large number of parameters,making them inapplicable on mobile devices.Against this backdrop,we propose a lightweight vehicle and pedestrian detection model,YOLOv8-TI(Traffic Information).A novel lightweight parameter-sharing SPG Detect detection head is designed to reduce the model's parameters and computational load.The Global Balanced Channel Path Aggregation Network(GBC-PAN)structure is proposed to balance the number of network channels and achieve bidirectional feature fusion from top-down and bottom-up directions through weighted connections across scales.Meanwhile,a dynamic non-monotonic focusing mechanism,represented by the Wise Loss function,is introduced to enhance the accuracy of predicted bounding boxes.Our experimental results reveal the YOLOv8-TI model maintains a high accuracy rate while reducing the parameters,flops,and model volume by 52.1％,58.0％and 54％respectively compared with those of YOLOv8n.A comparative analysis with other lightweight object detection algorithms verifies the effectiveness and superiority of our method.YOLOv8-TI is put on Android mobile devices and tested on Honor 20 fps and Honor 80GT,achieving frame rates of 24 and 31 FPS respectively,meeting real-time requirements.It is set to accomplish traffic information detection tasks when applied on autonomous driving vehicles.
关键词
深度学习/车辆行人检测/参数共享/轻量化Key words
deep learning/car and pedestrian detection/sharing Parameter/light weight引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据