密集交通场景的目标检测算法研究-《计算机技术与发展》

文章信息/Info

Author(s):: LI Xuan; LI Jing; WANG Hai-yan; School of Electronic Information Engineering,Shenyang Aerospace University,Shenyang 110136,China

Keywords:: dense object; regression loss function; matching degree; position information; YOLOv3; object detection

摘要:: 在现实的交通场景中,行人和车辆经常聚集在一起,形成相互遮挡的现象,给交通场景的目标检测带来了极大的挑战。针对交通场景中目标密集,位置接近造成的目标漏检、同一检测框中包含多个目标的问题,提出一种针对性遮挡回归损失函数 Occlusion Loss。 Occlusion Loss 有两个作用:一是指导神经网络学习检测框和真实框匹配程度得到更为准确的位置信息;二是在学习到位置信息后尽可能减少一个检测框有多个被检测目标的情况。将提出的 Occlusion Loss 应用到YOLOv3 目标检测算法上,经过实验证明改进后的 YOLOv3 在密集的交通场景中有更准确的检测结果,能够有效防止目标漏检现象,定位更加准确,具有很强的鲁棒性。在重新划分的交通场景数据集 KITTI 中准确率和召回率均有所提高,平均准确率达到 92.67% ,优于其他目标检测算法。

Abstract:: In the actual traffic scene,pedestrians and vehicles often gather together to form a mutual occlusion phenomenon,which brings great challenges to the object detection in traffic scenes. Aiming at the problem that the object is dense in the traffic scene, the object is missed, and the same detection frame contains multiple objects, we propose a targeted occlusion regression loss function Occlusion Loss which has two functions. The first is to guide the neural network to learn the matching degree of detection box and the groundtruths for more accurate position information. The second is to reduce the number of detected objects in one detection box as much as possible after learning the position information. The proposed occlusion loss is applied to the YOLOv3 object detection algorithm. It is proved by experiments that the improved YOLOv3 has more accurate detection results in dense traffic scenarios,which can effectively prevent object miss detection,with more accurate positioning and stronger robustness. In the re-divided traffic scene dataset KITTI,the accuracy and recall rate are improved,and the average accuracy rate is 92.67% ,which is better than other object detection algorithms.