基于深度学习的视频船舶目标追踪模型

Video Ship Target Tracking Model Based on Deep Learning

摘要: 船舶追踪是保证水上船舶交通安全的重要技术之一，船舶目标追踪模型可以对船舶的轨迹建模从而预测船只位置，监控船只交通安全，然而在船舶在交汇情境下，由于船舶被部分遮挡或处于复杂背景中，仅依赖视觉特征进行船舶追踪可能会导致追踪边界的不准确定位等问题。此外，当船只完全被遮挡并再次出现时，传统ReID方法往往不能为这些船只正确分配轨迹。为了解决这些问题，本文首先提出了一种融合光流场特征的单目标船舶追踪模型。该模型通过将视觉追踪特征与光流特征相结合，实现了更精确的目标边界定位。接着，通过结合Yolov5目标检测模型和经过改进的单目标追踪模型，实现了一种判别式多目标追踪模型，并采用更有效的匹配方法，一定程度上缓解了目标消失再出现id丢失的问题。实验表明，融合光流网络的目标追踪模型能更好地定位目标，而本文的联合多目标追踪框架在目标定位和ID分配方面均优于DeepSort、ByteTrack等追踪模型。

Abstract: Ship tracking is one of the crucial technologies for ensuring the safety of maritime traffic. Ship tar-get tracking models can model the trajectory of ships to predict their positions and monitor mari-time traffic safety. However, in situations where ships intersect, issues such as partial ship obstruc-tions or complex backgrounds can lead to inaccurate tracking boundary positioning when relying solely on visual features. Additionally, traditional discriminative tracking models often struggle to correctly assign identities to ships when they reappear after being completely obscured. To address these challenges, this paper first introduces a single-object ship tracking model that integrates op-tical flow field features. This model combines visual tracking features with optical flow features, re-sulting in more precise target boundary localization. Subsequently, a joint multi-object tracking framework is implemented by combining the Yolov5 object detection model with an improved sin-gle-object tracking model. This framework also utilizes more effective matching methods, partially alleviating the issue of losing IDs when targets reappear. Experimental results demonstrate that the target tracking model incorporating optical flow performs better in target localization. Fur-thermore, the joint multi-object tracking framework presented in this paper outperforms tracking models like DeepSort and ByteTrack in both target localization and ID assignment.