基于果蝇视觉神经网络的单目标运动检测与跟踪

Fly visual neural network based single target motion detection and tracking

运动目标跟踪因对实时输入信息的处理、跟踪模型的性能有较高要求,使得对应用性较强的目标跟踪模型的构建仍然是较为活跃的研究重点。针对单目标跟踪问题,基于果蝇视觉信息处理机制及目标跟踪具有的固有特征,本文提出不同于现有目标跟踪模型的新型模型。模型设计中,基于果蝇视觉信息处理机制建立改进型前馈果蝇视觉神经网络,进而借助其输出的运动方向量矩阵及运动目标的固有运动特性,构建运动目标的运动方向检测以及位置、速度、偏航角估计模型,由此获得计算复杂度由输入图像的分辨率确定的前馈果蝇视觉目标跟踪模型。比较性的实验表明,相较于经典的和基于深度学习的目标跟踪模型,所获新型目标跟踪模型在多种指标下具有实时处理能力强、跟踪效果好且有较好应用潜力的优点,为目标跟踪研究提供了又一新的解决方案。

Due to the high requirements of real-time input information processing and target tracking performance,the construction of target tracking models with strong applicability is still a relatively active research focus.To address the single target tracking problem,a novel target tracking model,which completely differs from existing target tracking models,is proposed based on the visual information processing mechanism of the fly and the intrinsic characteristics of the target tracking problem.In the design of the model,a feedforward fly visual target tracking neural network is developed,in which the computational complexity is decided by the input image’s resolution.Herein,on the one hand,an improved feedforward fly visual neural network model is acquired based on the fly visual information processing mechanism,on the other hand a computational model is constructed to perform the attitude estimation of a moving object,which involves the target’s motion direction detection,velocity estimation,and yaw angle estimation.Comparative experiments have validated that,compared to classical and deep learning-based object tracking models,the new model has significant advantages over the compared models based on the evaluation indicators.It exhibits the strong abilities of real-time processing,target tracking performance,and good application potential.This provides another new scheme to handling object tracking problems.