基于改进SSD的机动车检测方法

Vehicle Detection Method Based on Improved SSD

论文基于SSD搭建了机动车检测框架。应用聚类方法对机动车数据集进行数据挖掘以得到更符合车辆尺寸的先验包围框。针对机动车数据集的正负样本不平衡问题,论文引入级联SSD的网络结构。第一级SSD挖掘正负样本,第二级SSD根据第一级SSD预处理的指导来过滤掉大量的负样本。同时在级联SSD之间加入融合特征层,以提高特征提取能力。为了验证该方法,论文在DETRAC数据集上评估了改进的SSD网络,取得了69.96%的检测精度,比SSD提高了13.07%。从实验结果可以看出该方法具有较好的通用性,适用于机动车检测任务。

This paper builds a vehicle detection framework based on SSD.In order to get more prior bounding boxes that are in line with vehicle size,clustering method is applied to mine data of vehicle datasets.In view of the imbalance of positive and negative samples in vehicle datasets,this paper introduces a cascaded SSD network structure.The first level SSD mines positive and negative samples,and the second level SSD filters out a large number of negative samples according to the guidance of the first level SSD.To improve the ability of feature extraction,a fusion feature is added between cascaded SSD.In order to verify this method,this paper evaluates the improved SSD network on DETRAC dataset,and achieves 69.96 of detection accuracy which is 13.07%higher than SSD.The experimental results show that the method has good generality and is suitable for vehicle detection.