基于DenseNet的无人机在线多目标跟踪算法

多目标跟踪任务的目的,是对图像序列中不同的目标设置不同的编号(ID),最终得到不同目标的运动轨迹。本文针对跟踪过程中目标ID极易变化的现象,提出了一种新的在线多目标跟踪算法。算法主要包含三个步骤:输入预处理、特征提取和数据关联。其中预处理步骤使用NMS算法对输入的检测结果进行筛选;特征提取步骤使用密集连接的特征提取网络对目标进行外观特征的提取,输出特征向量矩阵;数据关联步骤则使用级联匹配的方式,依据目标的位置信息和外观特征信息为其分配各自的ID。此外,该文还整理了一个具有挑战性的无人机场景下的多目标跟踪测试集。实验结果表明,该方法有效地减少了错误的目标ID变化,提高了多目标跟踪算法面对复杂场景时的精度,并保持较快的运行速度。

Online Multi-Object Tracking Algorithm with Dense Connection Network

The main purpose of the multi-target tracking task is to set different identity numbers(ID) for different objects in an image sequence, and finally get the motion trajectories of different targets. Aiming at the phenomenon that the target ID is easy to change in the tracking process, this paper proposes a new online multi-target tracking algorithm. The algorithm consists of three steps: input preprocessing, feature extraction and data association. In the preprocessing step, the NMS algorithm is used to filter the input detection results;in the feature extraction step, a dense connection feature extraction network is used to extract the appearance features of the target, and the feature vector matrix is output;in the data association step, the corresponding ID is assigned according to the location information and appearance feature information of the target. On the other hand, we propose challenging multiple people tracking test set for UAV scenarios. The experimental results show that this method can effectively reduce the error of target ID changes, improve the accuracy of multi-target tracking algorithms in complex scenes, and maintain a fast running speed.