基于多级学习的长尾分布下交通多目标检测

Multi-level Learning-based Multi-objective Detection of TrafficUnder Long-tail Distribution

现有检测算法在复杂交通环境下受到长尾分布的影响,存在各类别性能难以平衡而导致精度下降的问题。因此,论文提出基于类别均衡的多级学习算法。首先在分类器上进行改进,采用多级分组分类的方式,实现较为平衡的类别学习过程。然后,引入基于多头注意力机制的特征分组方式,完成不同粒度语义信息的融合和特征提取。最后,为缓解组间的样本不平衡,构造了Logit联合调整方式,对两级类别输出进行调整。实验证明,论文提出的算法能有效缓解交通场景下的类别不平衡,提高了目标检测的准确性和鲁棒性。

Existing detection algorithms are affected by the long-tail distribution in complex traffic environments,and suffer from the problem of accuracy degradation due to the difficulty of balancing the performance of each category.Therefore,this paper proposes a multi-stage learning algorithm based on category balancing.Firstly,the classifier is improved by using multi-level group⁃ing classification to achieve a more balanced category learning process.Then,a feature grouping approach based on a multi-headed attention mechanism is introduced to complete the fusion and feature extraction of semantic information with different granularity.Fi⁃nally,to alleviate the sample imbalance between groups,the Logit joint adjustment method is constructed to adjust the two-level category output.Experiments prove that the algorithm proposed in this paper can effectively alleviate the category imbalance in traf⁃fic scenes and improve the accuracy and robustness of object detection.