【摘要】 行人检测技术因为其在智能安防领域具有广泛的应用前景而成为计算机视觉领域的研究热点之一,在简单的应用场景取得了一定的研究成果。但是实际监控场景包含行人所处环境复杂(例如物体遮挡、光照变化等),有虚警目标等问题,任何一种方法都很难单独保证在复杂场景下即不出现漏检又不出现虚警目标。本文针对这些问题,研究了深度学习中的YOLO-v3(You Only Look Once version3)方法和基于GMM(Gaussian of Mixture Models,混合高斯模型)的运动目标检测算法。针对视频行人检测这个特定的应用,本文将YOLO-v3与GMM建模方法相结合,充分利用YOLO能够准确检测出具有训练数据中目标特征的目标的优点,结合GMM算法弥补了YOLO无法利用目标运动信息的缺陷。实验结果表明,该方法相较于门限设置为0. 01的YOLO检测方法,在保证召回率在可接受范围内降低的同时,检测准确率提高了约3. 3倍,平均交并比提高了约3. 7倍;相较于门限设置为0. 3的YOLO检测方法,在保证准确率和平均交并比在可接受范围内降低的同时,召回率提高了约1. 9倍。更多还原

【Abstract】 Pedestrian detection technology has become one of the research hotspots in the field of computer vision because of its wide application prospects in the field of intelligent security. It has achieved certain research results in simple application scenarios. However,the actual monitoring scene includes the complex environment of the pedestrian(such as object occlusion,illumination changes,etc.),and there are problems such as false alarm targets. It is difficult to ensure that all targets are detected in a complicated scene. And there is no false alarm goal. In view of the above problems,YOLO-v3(You Only Look Once version 3) method and the moving object detection algorithm based on GMM(Gaussian of Mixture Models) are studied. For the specific application of video moving target detection,YOLO-v3 in combination with GMM modeling method makes full use of YOLO’s ability to accurately detect targets with target features in training data,which combines GMM algorithm to make up for YOLO’s inability to use targets. The experimental results show that compared with the YOLO detection method with the threshold setting to 0. 01,the detection accuracy is improved by 3. 3 times and the average cross-section ratio is increased by about 3. 7 times while ensuring that the recall rate is reduced within an acceptable range.Compared to the YOLO detection method with a threshold setting to 0. 3,the recall rate is increased by about 1. 9 times while ensuring that the accuracy and average cross-over ratio are reduced within an acceptable range.更多还原