改进YOLOv7+Bytetrack的小目标检测与追踪

Improved Small Target Detection and Tracking with YOLOv7+Bytetrack

近年来,目标检测技术已经相当成熟,但小目标检测一直是目标检测领域的一大挑战。为了解决这一问题,设计一种名为MFF-YOLOv7的小目标检测算法,该算法旨在提高小目标检测的准确率。设计级联双向特征金字塔KBiFPN,以及联合提出的多级感受野特征聚合模块MFA,来聚合浅层特征并增强特征的信息表达能力。为了解决小目标漏检问题,设计了新的解耦头和新的注意力机制。新的解耦头对小目标的检测能力更强,新的注意力机制可以重点关注感兴趣的小目标区域。引入了一种新的损失函数ECIOU,旨在加快模型的收敛速度。为了验证模型的性能,分别在三个小目标数据集上进行了实验。实验结果表明,MFF-YOLOv7算法提高了检测精度。同时,使用多目标追踪Bytetrack算法在MOT17和VisDrone2019-MOT两个多目标追踪数据集上对新模型进行了验证,进一步证明了其有效性。此外,MFF-YOLOv7算法在动态视频追踪中表现出了良好的性能。

In recent years,although target detection techniques have become quite mature,the detection of small targets has been a major challenge in the field of target detection.In order to solve this problem,a small target detection algorithm called MFF-YOLOv7 is developed.The algorithm aims to improve the accuracy of small target detection.Firstly,the designed cascaded bidirectional feature pyramid KBiFPN and the proposed multi-level sensory field feature aggregation module MFA are used to aggregate the shallow features and improve the information expression of the features to improve the accuracy of small target detection.To solve the problem of small target miss detection,a new decoupling head and a new attention mechanism are designed.This detection head is more capable of detecting small targets and can focus on the small target region of interest through the attention mechanism,thus reducing the leakage detection.Finally,a new loss function,ECIOU,is introduced to speed up the convergence of the model.To validate the performance of the model,experiments are performed on three small target datasets.The experimental results show that the MFF-YOLOv7 algorithm achieves higher detection accuracy.Meanwhile,the effectiveness of the new model is further demonstrated by the multitarget tracking Bytetrack algorithm on two multi-target tracking datasets,MOT17 and VisDrone 2019-MOT.MFF-YOLOv7 is also able to show good performance in dynamic video tracking.