基于多特征信息融合的目标检测方法研究

TARGET DETECTION METHOD BASED ON MULTI-FEATURE INFORMATION FUSION

针对目标检测中小目标漏检、准确率较低和容易受到相似目标干扰等问题,在SSD(Single Shot MultiBox Detector)模型基础上,提出一种融合低层手工特征和深层网络特征的目标检测方法。通过对图像提取融合多种目标特征,获取目标大概位置和类别。基于非极大值抑制(NMS)将交并比小于0.7的窗口消除,解决行人部分重叠与小目标的漏检问题,提升目标检测的准确性和目标检测模型的泛化能力。该模型在VOC2007公开数据集上的平均检测精度较SSD算法提升了4%,NMS机制的加入有效提升了目标检测速度和稳定性。

Aiming at the problems of small target missed detection,low accuracy and vulnerable to similar target interference,based on single shot multi-Box detector(SSD)model,we propose a target detection method combining low-level manual features and method for deep network features.The variety of target features were integrated through the image extraction,and the approximate location and category of the target were obtained.We eliminated the window with less than 0.7 ratio based on non-maximum value suppression(NMS),which solved the problem of pedestrian overlap and small target omission,and improved the accuracy of target detection and generalization of target detection model.The average detection accuracy of our model on VOC2007 public dataset is 4%higher than that of SSD algorithm,and NMS mechanism effectively improves the speed and stability of target detection.