基于候选区域选择及深度网络模型的骑车人识别
发布时间：2018-10-05 21:48
【摘要】：基于骑车人目标识别的骑车人保护系统是保护道路环境中骑车人的重要手段。该文提出了骑车人目标的候选区域选择方法,并结合基于深度卷积神经网络的目标分类与定位方法,实现了骑车人目标的有效识别。候选区域选择方法可分为3部分:骑车人共有显著性区域检测、基于冗余策略的候选区域生成和基于车载视觉几何约束的候选区域选择。在公开的骑车人数据库上进行的对比试验表明:相对于现有的目标候选区域选择及目标识别方法,该方法显著提升了骑车人目标的识别率及识别精度,进而验证了该方法的有效性。
[Abstract]:Biker protection system based on target recognition is an important method to protect cyclists in road environment. In this paper, a candidate region selection method for cyclists is proposed, and the method of target classification and location based on deep convolution neural network is combined to realize the effective recognition of cyclists. The candidate region selection method can be divided into three parts: rider common significant region detection, candidate region generation based on redundancy strategy and candidate region selection based on vehicle vision geometric constraints. A comparative experiment conducted on the open cyclists database shows that compared with the existing methods of target candidate selection and target recognition, this method has significantly improved the recognition rate and accuracy of cyclists' targets. The validity of the method is verified.
【作者单位】： 清华大学汽车安全与节能国家重点实验室;北京航空航天大学软件学院;
【基金】：国家自然科学基金资助项目(51605245) 戴姆勒-清华大学联合项目
【分类号】：TP391.41
本文编号：2254989
[Abstract]:Biker protection system based on target recognition is an important method to protect cyclists in road environment. In this paper, a candidate region selection method for cyclists is proposed, and the method of target classification and location based on deep convolution neural network is combined to realize the effective recognition of cyclists. The candidate region selection method can be divided into three parts: rider common significant region detection, candidate region generation based on redundancy strategy and candidate region selection based on vehicle vision geometric constraints. A comparative experiment conducted on the open cyclists database shows that compared with the existing methods of target candidate selection and target recognition, this method has significantly improved the recognition rate and accuracy of cyclists' targets. The validity of the method is verified.
【作者单位】： 清华大学汽车安全与节能国家重点实验室;北京航空航天大学软件学院;
【基金】：国家自然科学基金资助项目(51605245) 戴姆勒-清华大学联合项目
【分类号】：TP391.41
【相似文献】
相关硕士学位论文 前1条
1 古杰;基于大规模仿真数据的骑车人碰撞行为研究[D];清华大学;2014年
，本文编号：2254989
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2254989.html
