**作者：**Yijun Qian∗, Lijun Yu∗, Wenhe Liu , and Alexander G. Hauptmann

作者单位：Language Technologies Institute, Carnegie Mellon University（卡内基梅隆大学）

发布时间：2020

发布期刊/会议：CVPRW

出版商：IEEE

**论文全称：**ELECTRICITY: An Efficient Multi-camera Vehicle TrackingSystem for Intelligent City

论文地址：

论文代码：

地位：在AI City 2020挑战赛 track 3 上以0.4585分获得第一名

个人理解

• 创新点：
  • 提出了聚合损失函数（aggregation loss）$loss_{agg}$
  • 利用 Vehicle Re-identification 模块输出的 $Dis_{N×M}$ 和 MOT输出的跟踪结果 $T$，提出三个规则来更新跟踪结果 $T$ ，最终输出 MTMC 结果
• **为什么：**参加AI City 2020挑战赛(Track 3，城市尺度多摄像头车辆跟踪)
• 怎么做：
  • 使用的是聚合损失函数（aggregation loss）$loss_{agg}$，而实在难例挖掘的前提下，将**交叉熵损失函数（cross entropy loss，$loss_{xe}$）和三重熵损失函数（triplet loss，$loss_{tr}$）**合并在一起
  • 提出了 ELECTRICITY 多摄像头多车辆跟踪架构，该架构由四个部分组成

一、摘要

City-scale multi-camera vehicle tracking is an important task in the intelligent city and traffic management. It is quite challenging with large scale variance, frequent occlusion and appearance variance caused by viewing perspective difference. In this paper, we propose ELECTRICITY, an efficient multi-camera vehicle tracking system with aggregation loss and fast multi-target cross-camera tracking strategy. The proposed system contains four main modules. Firstly, we extract tracklets under single camera view through object detection and multi-object tracking modules which shared the detection features. After that, we match the generated tracklets through a multi-camera re-identification module. Finally, we eliminate isolated tracklets and synchronize tracking ids according to the re-identification results. The proposed system wins the first place in the City-Scale Multi-Camera Vehicle Tracking of AI City 2020 Challenge (Track 3)1 with a score of 0.4585.

• 城市层面的多摄像头车辆跟踪任务由于由于视点差异引起的大尺度方差、频繁的遮挡和外观差异是非常具有挑战性的