毕业生知识库

多目标跟踪是当前计算机视觉领域的热门研究方向之一，被广泛应用于运动分析、机器人视觉导航、智能视频监控以及自动驾驶辅助系统中。随着互联网时代的到来，大量监控相机遍布于国内外各大公共场所，每一秒都能产生海量的视频数据。面对海量的视频数据，如何高效检索和跟踪感兴趣目标成为亟待解决的难题之一。多目标跟踪技术经过了多年的研究与发展，已经取得诸多成果。但在现实世界中，被跟踪目标往往处于复杂场景中，比如目标频繁遮挡、光照变化、背景干扰和相机运动等，且跟踪目标数量较多，使得现有多目标跟踪算法难以同时满足高稳定和实时性的需求，难以在实际场景中进一步应用。

在跟踪视频序列中，连续两帧之间相应目标表观变化幅度较小，因此表观模型构建的好坏是影响跟踪过程的稳定性的一个重要因素。近年来，深度学习技术蓬勃发展，在特征提取方面表现出极大的潜力。针对复杂场景下目标跟踪算法的各种难点问题，本文对于现有算法进行进一步研究，将深度神经网络提取的目标表观特征和运动特征纳入多目标跟踪算法中数据关联部分，主要工作如下：

（1）针对目前计算机视觉中检测与跟踪出现的断层的问题，本文实现了一种面向监控场景的检测跟踪一体化的在线多目标跟踪算法。该算法对当前帧图像检测出目标位置并提取特征，将具有更强表达能力的深度特征应用于跟踪中，在一些目标相对稀疏的场景中可以实现实时跟踪；

（2）针对目前跟踪算法中采用一阶运动模型或者二阶运动模型，在目标速度变化较大时而导致目标预测位置不准确的问题，本文提出一种基于检测跟踪相关联的多目标跟踪算法。该算法利用两帧图像在特征图层面上进行相关，预测目标的位置，实现更加稳定的跟踪；

最后，本文进行一系列的实验对算法加以评估。实验结果表明，与一些主流算法相比较，本文方法取得了不错的跟踪效果，在部分指标上取得了领先。

Multi-Object Tracking(MOT) is one of the most popular research directions in the field of computer vision. It is widely used in sport analysis,robot vision navigation, intelligent video surveillance and autonomous driving assistant systems.With the advent of the Internet Age, quantities of monitoring cameras are applied among most major public places at home and abroad, generating massive amounts of video data every second. Faced with massive video data, how to implement efficient retrieval and object tracking has become one of the difficult problems to be solved. After years of research and development, multi-target tracking technology has achieved many goals. Nevertheless, in the real world, the tracked targets are often in complex scenes such as frequent target occlusion, illumination changes, background interference and camera motion, the number of tracking targets makes it difficult for existing multi-target tracking algorithms to simultaneously satisfy high stability and real-time needs. And it is hardly to further apply in actual scenarios.

In the tracking video sequences, the appearance of the corresponding target between two consecutive frames varies slowly, therefore, the construction of the appearance model is a significant factor for the stability of tracking process. In recent years, the deep learning technology has flourished and has shown great potential in learning discriminative visual features. To tackle these difficulties of multi-target tracking algorithms in complex scenes, this thesis further considers existing algorithms, and incorporates appearance feature and motion feature extracted by deep neural networks into data association of multi-target algorithm. The main work is as follows:

(1)For the split of detection and tracking in computer vision, we implement an online multi-target tracking algorithm that integrates detection and tracking for surveillance scenarios. The algorithm detects the location of targets and extract features in the current frame, and  applies deep features with stronger expressive power to tracking. Real-time tracking can be realized in some scenaros where targets are relatively less dense;

(2)This dissertation proposes a multi-target tracking algorithm based on correlation of detection and tracking to tackle the problem that current tracking algorithms use first-order motion model or second-order motion model predict the position unreliably when a target's speed varies violently. The algorithm uses two frames of images to correlate at the feature map level, predicting the position of the target, and achieving more stable tracking;

Finally, this thesis conducts a series of experiments for evaluation. The results show that this method perform favourably against several state-of-the-art methods and achieve a leading position in some factors as well.