基于SSD的不平衡样本车辆检测与识别  

Vehicle Detection and Recognition in Unbalanced Samples Based on SSD

作　　者：马浩良 谢林柏 MA Hao-liang;XIE Lin-bo(School of Internet of Things Engineering,Jiangnan University,Wuxi 214122,China)

机构地区：[1]江南大学物联网工程学院

出　　处：《计算机技术与发展》2019年第12期135-140,共6页Computer Technology and Development

基　　金：国家自然科学基金(61374047,60973095);江苏省博士后科研资助计划(1601085C)

摘　　要：为了实现在复杂环境,车辆样本不平衡情况下的实时车辆检测与识别,基于SSD算法搭建了车辆检测与识别的框架。针对车辆数据存在车型难易样本不均衡以及SSD方法存在的正负样本不平衡问题,在SSD引入改进的损失函数来挖掘难易样本,通过提高难样本的学习比例来更好地识别样本较少的车辆类型。引入SSD级联的网络结构,在第一级SSD挖掘正负样本,在第二级SSD根据第一级SSD的指导过滤掉大量的负样本。构建了拥有7480幅图像,包含4种车辆类型的数据集对该方法进行验证。实验结果表明,基于改进SSD的方法提高了少样本车辆类型的准确率,使整体检测精度取得了90.0%的准确率。针对不均衡样本的车辆数据集有较好的通用性,适用于车辆检测与识别任务。In order to realize real-time vehicle detection and recognition in a complex environment with unbalanced vehicle samples,a framework of vehicle detection and recognition is built based on SSD algorithm.Aiming at the imbalance of difficult and easy samples in vehicle data and the imbalance of positive and negative samples in SSD method,an improved loss function to mine difficult and easy samples in SSD is introduced to identify vehicle types with fewer samples better by increasing the learning proportion of difficult samples.The cascade SSD network structure is introduced to mine positive and negative samples in the first-level SSD and filter out a large number of negative samples in the second-level SSD according to the guidance of the first-level SSD.A data set with 7480 images and 4 vehicle types is constructed to verify this method.The experiment shows that the improved method based on SSD improves the accuracy of vehicle types with fewer samples and the overall detection accuracy achieves 90.0%.This method has excellent generality for vehicle data sets with unbalanced samples and is suitable for vehicle detection and recognition.

关 键 词：车辆检测与识别 SSD 样本不平衡 难易样本挖掘 正负样本挖掘 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]