面向复杂自然环境的红外目标跟踪数据集构建与评测

Construction and Benchmark Evaluation of an Infrared Target Tracking Dataset for Complex Natural Environments

摘要:

红外目标检测跟踪是智能感知、视频监控和自动驾驶中的关键技术,作为计算机视觉领域的重要分支,在恶劣环境下的感知任务中面临巨大挑战。然而,现有红外数据集大多面向通用场景,存在特定目标类别覆盖不足、背景环境单一以及复杂干扰因素缺失等问题,难以满足复杂动态场景下算法鲁棒性评估的需求。为此,本文构建了面向复杂自然环境的红外目标数据集WAR-TIR(Complex Environment Infrared Tracking Dataset)。该数据集采用实地采集、公开数据筛选与物理建模仿真相结合的方式构建,包含80条视频序列,涵盖车辆、无人机及其他移动目标等8类典型对象,覆盖田野、荒漠、山地及开放场地等多样化背景,并重点引入了重度雾霾、强光干扰、动态模糊及热辐射伪影等6类典型环境干扰。本文系统分析了数据集的构建流程、标注规范及统计特性,并选取KCF、SiamFC、SiamBAN、ATOM、PrDiMP和TransT等代表性算法建立了基准测试。实验结果表明,当前性能最优的TransT算法在WAR-TIR上的成功率和精确度仅分别为0.642和0.695。这表明该数据集具有较高的挑战性,能够有效评估跟踪算法在复杂环境下的抗干扰能力与泛化性能。

Abstract:

Infrared target detection and tracking serve as a pivotal technology for intelligent sensing, video surveillance, and autonomous driving. As a crucial branch of computer vision, it faces significant challenges in perception tasks under adverse environmental conditions. However, prevailing infrared datasets predominantly cater to generic scenarios, suffering from limited coverage of specific target categories, monotonous background environments, and a lack of complex interference factors. These limitations hinder the robust evaluation of algorithms in complex dynamic scenarios.This paper presents WAR-TIR (War Infrared Target Tracking Dataset), a dedicated infrared military target tracking dataset for complex battlefield scenarios. The dataset is constructed through a combination of real-world infrared acquisition, publicly available data collection, and simulation-based generation. WAR-TIR contains various representative military targets, including tanks, armored vehicles, fighter aircraft, helicopters, and unmanned aerial vehicles, under diverse battlefield backgrounds such as fields, deserts, mountains, and shooting ranges. In addition, several challenging battlefield interference factors, including smoke, flames, explosions, and point-source decoys, are incorporated to enhance dataset complexity. The proposed dataset consists of 80 video sequences and approximately 31K annotated frames, covering 8 military target categories and 6 interference categories. Comprehensive analyses of dataset construction, annotation strategy, statistical characteristics, and challenge attributes are conducted. Furthermore, benchmark evaluations are established using six representative tracking algorithms, including KCF, SiamFC, SiamBAN, ATOM, PrDiMP, and TransT. Experimental results show that the best-performing tracker, TransT, achieves only 0.642 success rate and 0.695 precision on WAR-TIR, indicating the significant challenges posed by the dataset. The proposed WAR-TIR dataset provides a valuable benchmark platform for evaluating and advancing infrared military target tracking algorithms in complex battlefield environments.