【摘要】 伴随海洋经济的不断发展，目标跟踪作为海洋安全的重要部分，需要兼顾高实时性与高准确性。针对多目标跟踪过程中常出现模糊匹配错误以及深度学习目标检测器推理速度慢的问题，提出了一种基于稀疏多假设匹配的轻量化红外多目标跟踪算法。首先，从定量角度分析费用矩阵的匹配过程，结合深度级联匹配算法与多假设跟踪算法，设计并实现了稀疏多假设匹配算法，在稀疏化费用矩阵、减少匹配过程计算量的同时，提高了费用矩阵匹配精度；其次，针对深度学习模型参数多、存在冗余参数的问题，采用层自适应剪枝算法对YOLOv8s模型进行剪枝，在不牺牲模型准确度的同时，减少模型的参数量与浮点运算量，使得YOLOv8s模型能够在更广泛的场景下部署。在视频序列跟踪实验中，MOTA指标较其他算法提高了0.2%～1.2%，证明稀疏多假设匹配算法提升了目标跟踪效果；在剪枝实验中，相比于原模型，剪枝后的YOLOv8s模型的平均精度(mAP@50)提升了0.1%，参数量下降为42.0%，浮点运算量下降为66.4%，实现了推理准确度与推理速度的提升。更多还原

【Abstract】 Objective Along with the continuous development of the marine economy, marine economic security is crucial.As an important research direction in marine intelligent sensing, ship multi-object tracking must consider high real-time and high accuracy. By quantitatively analyzing the matching process of the cost matrix, the sparse multihypothesis(SMH) matching algorithm is designed by combining the deep cascade matching(DCM) algorithm with the multiple hypothesis tracking(MHT), which improves the matching accuracy of the cost matrices while sparsifying the cost matrices and reducing the amount of computation in the matching process. Aiming at the problem of redundant parameters of the deep learning model, the YOLOv8s model is pruned using the layeradaptive magnitude-based pruning(LAMP) algorithm to reduce the number of parameters and floating-point operations(FLOPs) of the model without reducing the accuracy of the model.Methods First, the YOLOv8s model is pruned to obtain the YOLOv8s-prune model. The YOLOv8s-prune model is utilized to detect the objects for each frame in a given video sequence and obtain the object detection results. Secondly, the detection results and trajectory prediction positions are grouped based on their pseudo-depth information, which provides the basis for judging standard matching and ambiguous matching. Next, the similarity between the corresponding groups of detection results and trajectory prediction locations is calculated using IoU to obtain the cost matrix group. The hypothesis tree and standard cost matrix group are constructed by analyzing the rows and columns of each cost matrix within the cost matrix group. The hypothesis trees represent several hypotheses composed of the trajectory prediction positions and detection results in ambiguous matching.The standard cost matrix group records the matching costs between the trajectory-predicted positions and detection results in standard matching. Finally, the optimal object tracking result can be obtained by solving the hypothesis trees and the standard cost matrix group. The problem of solving the hypothesis tree is transformed into solving the maximum weighted independent set(MWIS) problem by calculating the trajectory score and solved by the Bron-Kerbosch algorithm; The standard cost matrix group is solved by the DCM algorithm from far to near based on the pseudo-depth information.Results and Discussions In comparisons with other trackers, SMH outperforms other object trackers in most metrics, which shows that SMH is more robust and effective(Tab.1). In the experiment on speed between MHT and SMH, SMH is faster with lower computational complexity(Tab.2). In the comparison experiments between YOLOv8s and YOLOv8s-prune, the number of YOLOv8s-prune parameters and the amount of FLOPs decreased dramatically. However, the accuracy of YOLOv8s-prune does not decrease a lot and even improved a little(Tab.3). In ablation experiments, SMH outperforms baseline and YOLOv8s-prune does not have an enormous impact on the accuracy of the tracker(Tab.4). In the tracking visualization between BoT-SORT and SMH, SMH can follow the original track number without assigning new track number when re-tracking a lost trajectory(Fig.10). In the response visualization of feature map, for small objects, YOLOv8s-prune has a weaker response to the object; For large objects, YOLOv8s-prune has a weaker response to the object boundary(Fig.11). In the channels visualization between YOLOv8s and YOLOv8s-prune, the results show that many channels in the large convolution kernel contribute weakly to the accuracy of the model. The modules with fewer channels are preserved(Fig.12).Conclusions A lightweight infrared multi-object tracking algorithm based on sparse multi-hypothesis matching is proposed to address the problems of ambiguous matching and the slow inference speed of deep learning-based object detectors. SMH combines the DCM algorithm with the MHT algorithm, outperforming existing multiobject tracking algorithms. Compared with YOLOv8s, the YOLOv8s-prune model obtained by the LAMP pruning algorithm has fewer parameters and floating-point operations. In the future, we plan to explore the temporal information in the video and combine it with the current spatial matching methods to improve the multiobject tracking effect.更多还原