融合全卷积网络的低空动态复杂场景多目标实时跟踪研究

Target detection and tracking is one of the criterions of UAV intelligence, low altitude multi target real-time detection and tracking is a typical problem in UAV vision monitoring field. Real time video monitoring for low altitude UVA is very difficult for various reasons, such as moving viewpoint, secene varied greatly, many moving targets with numbers and categories unknown, there have been many unknown target out and in the sight, little difference within the same category object, mutual occlusion, etc. This project starts with setting up target datasets for low-level dynamic complex scenes, which proposes a multi-objective tracking framework for transfer-regression, and simultaneously studies multi-objective precise detection algorithms that infer tens of thousands of subclasses using data acquisition height information fusion full convolutional networks. When the targets decrease or increase, the trajectory of the return-to-regression skip frame is checked. By exploring the mechanisms and rules of mutual influence between the targets and blocking each other's movement information, the corresponding model is established and the result is fed back to the precise real-time detection trajectory. It will achieve accurate and stable multi-target tracking in low altitude complex situations. Farther more, this project establish tracklets association method to reduce the tracking drift to ensure the tracking process last for longer time. The proposed research will promote the development of UAV low altitude target detection and tracking technology, and improve the level of UAV intelligence, which has important theoretical value and practical application prospects.

目标检测与跟踪是无人机智能化的标准之一，多目标实时检测与跟踪是无人机低空视觉监控领域的经典难题。无人机低空实时视频监控面临视点移动、场景变化大、运动目标数量多类别多且数量未知、不断有目标离开视线新目标进入视野、类内目标类差异较小、目标间运动相互影响、目标间相互遮挡等问题。本项目拟从建立低空动态复杂场景目标数据集入手，提出转移回归多目标跟踪框架，研究利用数据采集高度信息融合全卷积网络加速推断上万个细类的多目标精确检测算法。当目标减少或增加时，启动转移回归跳帧核查轨迹，通过探索目标间相互影响、相互遮挡运动信息的机理和规律，建立对应模型并将结果反馈给精确实时检测轨迹，实现低空复杂情况下多目标准确稳定跟踪。最后形成一套轨迹关联方法，降低跟踪出现漂移的现象保证长时间持续跟踪。研究成果将推动无人机低空多目标跟踪方法的发展与技术突破，提升无人机智能化水平，具有重要的理论价值和实际应用前景。

实时目标检测与跟踪是实现智能飞行的基本能力之一，研究无人机低空目标检测与跟踪具有重要价值。本项目首先研究了交通场景多目标检测、目标间相互影响及其相互遮挡建模、对模糊多目标跟踪。其次，形成一套轨迹关联方法，降低跟踪出现漂移的现象保证长时间持续跟踪。最后，利用稀疏光学的组合算法对目标检测跟踪后的异常轨迹，通过均值漂移聚类实现非监督学习低空车辆异常行为监控。取得的进展如下：.（1）通过加入辅助的非跟踪训练数据集对现有跟踪数据集进行扩充，来训练低空对地目标检测器。对于数量庞大的视频样本集，采用计算开销较小的启发式方法单遍法聚类，对有限样本信息对预训练模型微调，抽取高层特征主要反映目标的语义特性，低层特征保存了更多细粒度的空间特性，对跟踪目标的精确检测更有效。车辆检测的均方根误差不断缩小达到3.6，F1达到0.947。.（2）研究对关联目标进行运动建模，利用空间邻近运动时序信息对目标间相互，相互遮挡时进行联合跟踪，最大限度地减少错误跟踪和恢复丢失数据的代价，提高跟踪效率和鲁棒性。.（3）对于连续状态下地面目标轨迹，研究选择让目标函数值下降的路线跳动，通过网络短时约束轨迹段度量学习，来提升轨迹段关联性。.（4）提出了一种新的无监督实时框架，以最大效率、准确的异常检测。此外，在道路交通监控场景中，检测速度也是一个关键因素，因为事故需要及时检测和响应。本项目解决了该问题，实现跟踪后对车辆行为进行包括碰撞、停滞、偏离轨迹等违规行为的理解。在Nvidia AI City Challenge数据集上实现实时检测异常，检测速度可达45.07 fps，S4-Score为0.936，属于本任务的综合指标领先水平。.研究成果将推动无人机低空多目标跟踪及行为监控的发展与技术突破，提升无人机智能化水平，具有重要的理论价值和实际应用前景。