| 基于YOLOv2模型的道路目标检测改进算法 |
| |
| 引用本文： | 宋建国,吴 岳.基于YOLOv2模型的道路目标检测改进算法[J].教育技术导刊,2019,18(12):126-129. |
| |
| 作者姓名： | 宋建国 吴 岳 |
| |
| 作者单位： | 山东科技大学 计算机科学与工程学院，山东 青岛 266590 |
| |
| 摘 要： | 针对传统道路目标检测算法推荐窗口冗余、鲁棒性差、复杂度较高的问题，提出基于YOLOv2模型的道路目标检测改进算法。相较于传统的HOG+SVM目标检测算法，YOLO模型优势在于提升了检测速度及准确度，更适用于实时目标检测。比较YOLO V3 与 YOLO V2算法，前者在构造神经网络模型时复杂度较高，故最终选择YOLO V2算法。针对原算法中选取Anchor Boxes时所采用的K-MEANS算法造成的目标物体框冗余问题，以及原算法对于不规则物体以及遮挡物体检测效果较差等问题，提出基于YOLO V2模型的一种改进方法，将K-MEANS算法改进为一种DA-DBSCAN算法，通过动态调整参数的方式大大减少了锚点框冗余问题。实验表明，改进后的模型准确率达到96.76%，召回率达到96.73%，检测帧数达到37帧/s，能够满足实时性要求。
|
| 关 键 词： | 目标检测算法 鲁棒性 深度学习 不规则 DA-DBSCAN 锚点框 |
| 收稿时间： | 2019-03-08 |
Improved Road Target Detection Algorithm Based on YOLOv2 Model |
| |
| SONG Jian-guo,WU Yue.Improved Road Target Detection Algorithm Based on YOLOv2 Model[J].Introduction of Educational Technology,2019,18(12):126-129. |
| |
| Authors: | SONG Jian-guo WU Yue |
| |
| Institution: | School of Computer Science and Engineering， Shandong University of Science and Technology， Qingdao 266590， China |
| |
| Abstract: | Aiming at the problem of traditional window target detection algorithm with recommended window redundancy， poor robustness and high complexity， an improved road target detection algorithm based on YOLOv2 model is proposed. Compared with the traditional HOG+SVM target detection algorithm， the YOLO model has the advantages of improved detection speed and improved accuracy， and is more suitable for real-time target detection. Comparing the YOLO V3 and YOLO V2 algorithms， the former has a higher complexity in constructing the neural network model， so the YOLO V2 algorithm is finally selected. Aiming at the problem of frame object redundancy caused by the K-MEANS algorithm used in the original algorithm when selecting Anchor Boxes， and the problem that the original algorithm has poor detection effect on irregular objects and occluded objects， it is proposed based on YOLO V2. An algorithm improvement method of the model improves the K-MEANS algorithm into a DA-DBSCAN algorithm. By dynamically adjusting the parameters， the anchor point redundancy problem is greatly reduced. Finally， the improved model accuracy rate reaches 96.76%， the recall rate reaches 96.73%， and the number of detection frames per second reaches 37 seconds per frame， which can meet the real-time performance. Requirements. |
| |
| Keywords: | target detection algorithm robustness deep learning irregularity DA-DBSCAN anchor box |
|
| 点击此处可从《教育技术导刊》浏览原始摘要信息 |
| 点击此处可从《教育技术导刊》下载免费的PDF全文 |
