Zhang Guoqing, Liu Jiandong, Zhao Yongxiang, Luo Wei, Mei Keyu, Wang Penggang, Song Yubin, Li Xiaoliang
North China Institute of Aerospace Engineering, Langfang, China.
Aerospace Remote Sensing Information Processing and Application Collaborative Innovation Center of Hebei Province, Langfang, China.
PLoS One. 2025 Jan 10;20(1):e0316933. doi: 10.1371/journal.pone.0316933. eCollection 2025.
As the global economy expands, waterway transportation has become increasingly crucial to the logistics sector. This growth presents both significant challenges and opportunities for enhancing the accuracy of ship detection and tracking through the application of artificial intelligence. This article introduces a multi-object tracking system designed for unmanned aerial vehicles (UAVs), utilizing the YOLOv7 and Deep SORT algorithms for detection and tracking, respectively. To mitigate the impact of limited ship data on model training, transfer learning techniques are employed to enhance the YOLOv7 model's performance. Additionally, the integration of the SimAM attention mechanism within the YOLOv7 detection model improves feature representation by emphasizing salient features and suppressing irrelevant information, thereby boosting detection capabilities. The inclusion of the partial convolution (PConv) module further enhances the detection of irregularly shaped or partially occluded targets. This module minimizes the influence of invalid regions during feature extraction, resulting in more accurate and stable features. The implementation of PConv not only improves detection accuracy and speed but also reduces the model's parameters and computational demands, making it more suitable for deployment on computationally constrained UAV platforms. Furthermore, to address issues of false negatives during clustering in the Deep SORT algorithm, the IOU metric is replaced with the DIOU metric at the matching stage. This adjustment enhances the matching of unlinked tracks with detected objects, reducing missed detections and improving the accuracy of target tracking. Compared to the original YOLOv7+Deep SORT model, which achieved an MOTA of 58.4% and an MOTP of 78.9%, the enhanced system achieves a MOTA of 65.3% and a MOTP of 81.9%. This represents an increase of 6.9% in MOTA and 3.0% in MOTP. After extensive evaluation and analysis, the system has demonstrated robust performance in ship monitoring scenarios, offering valuable insights and serving as a critical reference for ship surveillance tasks.
随着全球经济的扩张，水路运输对物流行业变得越来越重要。这种增长为通过应用人工智能提高船舶检测和跟踪的准确性带来了重大挑战和机遇。本文介绍了一种为无人机设计的多目标跟踪系统，分别利用YOLOv7和Deep SORT算法进行检测和跟踪。为了减轻有限船舶数据对模型训练的影响，采用迁移学习技术来提高YOLOv7模型的性能。此外，在YOLOv7检测模型中集成SimAM注意力机制，通过强调显著特征和抑制无关信息来改善特征表示，从而提高检测能力。包含部分卷积（PConv）模块进一步增强了对形状不规则或部分遮挡目标的检测。该模块在特征提取过程中最小化无效区域的影响，从而产生更准确和稳定的特征。PConv的实现不仅提高了检测精度和速度，还减少了模型的参数和计算需求，使其更适合在计算能力受限的无人机平台上部署。此外，为了解决Deep SORT算法聚类过程中的漏检问题，在匹配阶段将IOU度量替换为DIOU度量。这种调整增强了未链接轨迹与检测到的物体的匹配，减少了漏检并提高了目标跟踪的准确性。与原始的YOLOv7+Deep SORT模型相比，其MOTA为58.4%，MOTP为78.9%，增强后的系统MOTA为65.3%，MOTP为81.9%。这意味着MOTA提高了6.9%，MOTP提高了3.0%。经过广泛的评估和分析，该系统在船舶监测场景中表现出强大的性能，提供了有价值的见解，并为船舶监视任务提供了重要参考。
