针对斑马鱼幼鱼的微小目标群体轨迹跟踪算法
Micro-Object Group Trajectory Tracking Algorithm for Zebrafish Larvae

针对斑马鱼幼鱼体积小，外观相似度高，运动模式不稳定等特征导致的目标丢失、目标身份匹配错误、轨迹混乱等问题，本文提出了一种基于卷积神经网络的多目标跟踪算法。算法分为目标检测和目标跟踪两部分。检测部分通过经典的检测算法YOLO (You Look Only Once) v5实现，跟踪部分，主要通过DeepSORT实现对斑马鱼幼鱼位置的跟踪与轨迹连接。针对幼鱼容易被漏检、错检等特点，在YOLOv5m的基础上改进了网络的检测层结构，提升对小目标的检测能力。同时融合双通道注意力机制(Convolutional Block Attention Module, CBAM)对网络对背景噪声的抗噪能力。结果表明，在小样本训练的情况下，本文提出的YOLOv5m-ss对斑马鱼幼鱼群体的检测精度@mAP (mean Average Precision)可达99.9%，相较原网络提升了9.4%。结合DeepSORT后的跟踪精度MOTA可达98.0%，在算法运行速度和精确率上都有一定优势。
In this paper, we propose a multi-object tracking algorithm based on convolutional neural networks aiming at the problems of small size, high appearance similarity, unstable motion pattern and other characteristics of zebrafish larvae, such as target loss, target identity matching error, and trajectory confusion. The algorithm is divided into two parts: object detection and object tracking. The detection part is realized by the classic detection algorithm YOLO (You Look Only Once) v5, and the tracking part is mainly realized by DeepSORT to track and connect the position of zebrafish larvae. Aiming at the characteristics of young fish being easy to be missed and misdetected, the detection layer structure of the network is improved on the basis of YOLOv5m to improve the detection ability of small targets. At the same time, the dual-channel Attention mechanism (Convolutional Block Attention Module, CBAM) is fused to improve the anti-noise ability of the network against background noise. The results show that in the case of small sample training, compared with the original YOLOv5 network, the detection accuracy @mAP (mean Average Precision) of YOLOv5m-ss proposed in this paper for zebrafish juvenile population can reach 99.4%. After combining DeepSORT, the tracking accuracy MOTA can reach 98.0%, which has certain advantages in algorithm running speed and accuracy.