本文为美国内华达大学拉斯维加斯分校（作者：Santosh Bhusal）的硕士论文，共72页。

本文的主要目的是研究现有的基于视觉的检测和跟踪算法在基于热图像视频监控中的性能。虽然基于颜色的监控已经得到广泛的研究，但这些技术不能在低照度、夜间或灯光变化和阴影限制其适用性的情况下使用。本论文的主要贡献有：（1）建立了一个新的彩色热数据集；（2）详细比较了不同的基于颜色的热数据检测和跟踪算法的性能；（3）提出了一种自适应神经网络来抑制误检测。

由于热视频监控的数据集不多，为了评价热图像中流行的基于颜色的检测和跟踪性能，我们收集了一个新的UNLV热彩色行人数据集。该数据集提供了人类在庭院中行走的俯视图，适用于无人机系统（UAS）等空中监视场景。研究了三种常用的热行人检测方案：1）类Haar特征，2）局部二值模式（LBP）和3）背景减法运动检测。A i）卡尔曼滤波预测器和iii）光流跟踪。结果表明，将Haar和LBP检测与50%重叠规则相结合，采用Kalman滤波器进行跟踪，可使检测的真阳性率（TPR）提高20%。然而，基于运动的方法在非运动摄像机场景中更能抑制假阳性。带LBP检测的Kalman滤波器是最有效的跟踪器，但光流能更好地抑制虚假噪声检测。本文还提出了一种基于“热图”的行人检测学习与表征技术，以及一种以目标为中心的无人机运动补偿方法。最后，提出了一种基于神经网络的误差反向传播自适应检测方法。自适应抑制方案能够成功地学习识别静态的错误检测，从而提高检测性能。

The main objective behind this thesis is to examine how existing vision-based detection and tracking algorithms perform in thermal imagery-based video surveillance. While color-based surveillance has been extensively studied, these techniques can not be used during low illumination, at night, or with lighting changes and shadows which limits their applicability. The main contributions in this thesis are (1) the creation of a new color-thermal dataset, (2) a detailed performance comparison of different color-based detection and tracking algorithms on thermal data and (3) the proposal of an adaptive neural network for false detection rejection. Since there are not many publicly available datasets for thermal-video surveillance, a new UNLV Thermal Color Pedestrian Dataset was collected to evaluate the performance of popular color-based detection and tracking in thermal images. The dataset provides an overhead view of humans walking through a courtyard and it appropriate for aerial surveillance scenarios such as unmanned aerial systems (UAS). Three popular detection schemes are studied for thermal pedestrian detection: 1) Haar-like features, 2) local binary pattern (LBP) and 3) background subtraction motion detection. A i) Kalman filter predictor and iii) optical flow are used for tracking. Results show that combining Haar and LBP detections with a 50% overlap rule and tracking using Kalman filters can improve the true positive rate (TPR) of detection by 20%. However, motion-based methods are better at rejecting false positive in non-moving camera scenarios. The Kalman filter with LBP detection is the most efficient tracker but optical flow better rejects false noise detections. This thesis also presents a technique for learning and characterizing pedestrian detections with ”heat maps”and an object-centric motion compensation method for UAS. Finally, an adaptive method to reject false detections using error back propagation using a neural network. The adaptive rejection scheme is able to successfully learn to identify static false detections for improved detection performance.

1.       引言

2. 文献回顾

3. 系统概述

4. 数据集

5. 行人检测

6. 行人跟踪

7. 活动分析

8. 边缘神经网络的抗误判特性

9. 结果与讨论

10. 结论与展望

更多精彩文章请关注公众号：