Sheng Wenshun, Shen Jiahui, Chen Qi, Huang Qiming
Pujiang Institute, Nanjing Tech University, Nanjing, Jiangsu, China.
PLoS One. 2025 Jun 4;20(6):e0322919. doi: 10.1371/journal.pone.0322919. eCollection 2025.
A real-time stable multi-target tracking method based on the enhanced You Only Look Once-v8 (YOLOv8) and the optimized Simple Online and Realtime Tracking with a Deep association metric (DeepSORT) for real-time stable multi-target tracking (S-YOFEO) is proposed to address the issue of target ID transformation and loss caused by the increase of practical background complexity. The complexity of the real-world context poses a great challenge to multi-target tracking systems. Changes due to weather or lighting conditions, as well as the presence of numerous visually similar objects, can lead to target ID switching and tracking loss, thus affecting the system's reliability. In addition, the unpredictability of pedestrian movement increases the difficulty of maintaining consistent and accurate tracking. For the purpose of further enhancing the processing capability of small-scale features, a small target detection head is first introduced to the detection layer of YOLOv8 in this paper with the aim of collecting more detailed information by increasing the detection resolution of YOLOv8 to ensure precise and fast detection. Secondly, the Omni-Scale Network (OSNet) feature extraction network is implemented to enable accurate and efficient fusion of the extracted complex and comparable feature information, taking into account the restricted computational power of DeepSORT's original feature extraction network. Again, addressing the limitations of traditional Kalman filtering in nonlinear motion trajectory prediction, a novel adaptive forgetting Kalman filter algorithm (FSA) is devised to enhance the precision of model prediction and the effectiveness of parameter updates to adjust to the uncertain movement speed and trajectory of pedestrians in real scenarios. Following that, an accurate and stable association matching process is obtained by substituting Efficient-Intersection over Union (EIOU) for Complete-Intersection over Union (CIOU) in DeepSORT to boost the convergence speed and matching effect during association matching. Last but not least, One-Shot Aggregation (OSA) is presented as the trajectory feature extractor to deal with the various noise interferences in complex scenes. OSA is highly sensitive to information of different scales, and its one-time aggregation property substantially decreases the computational overhead of the model. According to the trial results, S-YOFEO has made some developments as its precision can reach 78.2% and its speed can reach 56.0 frames per second (FPS), which fully meets the demand for efficient and accurate tracking in actual complex traffic environments. Through this significant increase in performance, S-YOFEO can contribute to the development of more reliable and efficient tracking systems, which will have a profound impact on a wide range of industries and promote intelligent transformation and upgrading.
为解决实际背景复杂性增加导致的目标ID转换和丢失问题，提出了一种基于增强型You Only Look Once-v8（YOLOv8）和优化的带深度关联度量的简单在线实时跟踪（DeepSORT）的实时稳定多目标跟踪方法（S-YOFEO）。现实世界环境的复杂性对多目标跟踪系统构成了巨大挑战。天气或光照条件的变化，以及大量视觉上相似物体的存在，可能导致目标ID切换和跟踪丢失，从而影响系统的可靠性。此外，行人运动的不可预测性增加了保持一致和准确跟踪的难度。为了进一步提高小尺度特征的处理能力，本文首先在YOLOv8的检测层引入了一个小目标检测头，旨在通过提高YOLOv8的检测分辨率来收集更详细的信息，以确保精确快速的检测。其次，考虑到DeepSORT原始特征提取网络的计算能力受限，实现了全尺度网络（OSNet）特征提取网络，以实现对提取的复杂且可比的特征信息进行准确高效的融合。再次，针对传统卡尔曼滤波在非线性运动轨迹预测方面的局限性，设计了一种新颖的自适应遗忘卡尔曼滤波算法（FSA），以提高模型预测的精度和参数更新的有效性，以适应实际场景中行人不确定的运动速度和轨迹。随后，通过在DeepSORT中用高效交并比（EIOU）替代完全交并比（CIOU），获得了准确稳定的关联匹配过程，以提高关联匹配过程中的收敛速度和匹配效果。最后但同样重要的是，提出了一次性聚合（OSA）作为轨迹特征提取器来处理复杂场景中的各种噪声干扰。OSA对不同尺度的信息高度敏感，其一次性聚合特性大大降低了模型的计算开销。根据试验结果，S-YOFEO取得了一定进展，其精度可达78.2%，速度可达每秒56.0帧（FPS），完全满足实际复杂交通环境中高效准确跟踪的需求。通过性能的显著提升，S-YOFEO可为更可靠、高效的跟踪系统的发展做出贡献，这将对广泛的行业产生深远影响，并推动智能转型和升级。
