Public security monitoring is a hot issue that the government and citizens pay close attention to. Multiobject tracking plays an important role in solving many problems for public security. Under crowded scenarios and emergency places, it is a challenging problem to predict and warn owing to the complexity of crowd intersection. There are still many deficiencies in the research of multiobject trajectory prediction, which mostly employ object detection and data association. Compared with the tremendous progress in object detection, data association still relied on hand-crafted constraints such as group, motion, and spatial proximity. Emergencies usually have the characteristics of mutation, target diversification, low illumination, or resolution, which makes multitarget tracking more difficult. In this paper, we harness the advance of the deep learning framework for data association in object tracking by jointly modeling pedestrian features. The proposed deep pedestrian tracking SSD-based model can pair and link pedestrian features in any two frames. The model was trained with open dataset, and the results, accuracy, and speed of the model were compared between normal and emergency or violent environment. The experimental results show that the tracking accuracy of mAP is higher than 95% both in normal and abnormal data sets and higher than that of the traditional detection algorithm. The detection speed of the normal data set is slightly higher than that of the abnormal data set. In general, the model has good tracking results and credibility for multitarget tracking in emergency environment. The research provides technical support for safety assurance and behavior monitoring in emergency environment.

中文翻译：

---

突发紧急情况或暴力环境中基于 SSD 的深度行人跟踪模型

治安监控是政府和市民关注的热点问题。多目标跟踪在解决公共安全的许多问题中发挥着重要作用。在拥挤的场景和紧急情况下，由于人群交叉的复杂性，预测和预警是一个具有挑战性的问题。多目标轨迹预测的研究还存在很多不足，多采用目标检测和数据关联。与目标检测的巨大进步相比，数据关联仍然依赖于手工制作的约束，例如组、运动和空间接近度。突发事件通常具有突变、目标多样化、低照度或分辨率等特点，这使得多目标跟踪更加困难。在本文中，我们通过联合建模行人特征，利用深度学习框架在对象跟踪中进行数据关联的进步。提出的基于深度行人跟踪 SSD 的模型可以配对和链接任意两帧中的行人特征。模型采用开放数据集训练，在正常环境和紧急或暴力环境下比较模型的结果、准确性和速度。实验结果表明，mAP在正常和异常数据集的跟踪精度均高于95%，高于传统检测算法。正常数据集的检测速度略高于异常数据集。总的来说，该模型对紧急环境下的多目标跟踪具有良好的跟踪效果和可信度。