基于一种卷积神经网络的目标快速跟踪技术

Track and detection for specified moving object based on Convolutional Neural Networks

摘要:

特定动态目标的快速检测及跟踪，是计算机视觉领域重要的课题。改变特征图在YOLOv3卷积神经网络中的选取位置，通过收集相关网络数据(类似模式分析、统计建模和计算学习视觉对象类别数据集合，即PASCAL Visual Object Classes数据集)构建自定义数据集合进行训练，使用面积的交并比完成辅助类别的联合，构建了能够实时检测特定目标在相关可视对象类检测数据集合上mAP@75达到47.41的检测器。联合卡尔曼滤波和匈牙利算法，通过将面积信息加入到匈牙利算法的代价矩阵中，改善了使用原方法产生大量ID切换(ID switch)的问题。该方法满足快速识别与跟踪的要求，在使用一张NVIDIA GeForce GTX 1060 6GB GPU条件下，平均速度能达到0.109 7 s/帧。

Abstract:

The detection and track of specified moving small object is an important subject in Computer Vision. By changing the position of the feature maps for fusion in the YOLOv3, building the custom database including three classes, and completing the combination of classes by using Intersection Over Union(IOU), a detector is created, which is able to detect the specified moving small object and makes mAP@75 reach 47.41 in the test customer's data set. Combining Kalman Filter and Hungarian method, and putting the scale information of predicted bounding box and ground bounding box, the detector can track the object and reduce the ID switch caused by camera's fast movement. The whole system's speed reaches up to 0.109 7 s/frame using one NVIDIA GeForce GTX 1060 6GB GPU.