摘要
为了提高牛场无人机目标跟踪算法的实时性和鲁棒性,试验以无人机跟踪牛只图像为研究对象,提出了一种基于残差累积模板的轻型孪生网络(siamese tracker with residual accumulation template, SiamRAT)目标跟踪算法,即采用轻量级卷积网络MobileNetV2为特征提取网络及以锚框比率变化为契机的模板更新机制,提高了算法的实时性;采用高置信度残差累积模板和多峰欧式距离检测模块来解决因相似牛只干扰而产生的跟踪漂移问题;最后将SiamRAT算法与SiamRPN++、SiamDW、DaSiamRPN、SiamRPN、ECO-HC算法在由无人机采集牧场牛只视频制作的测试数据集和VOT2018数据集中相同属性视频构成的测试数据集上,以平均精确度、鲁棒性及帧率(frames per second, FPS)为指标进行性能比较,并分析改进模块(包括残差累积模板、高置信度更新和峰值距离检测3个模块的改进)对SiamRAT算法的贡献。结果表明:与SiamRPN++、SiamDW、DaSiamRPN、SiamRPN、ECO-HC算法相比,SiamRAT算法鲁棒性最优,平均精确度稍有下降,但仍处于所有算法的第二位;FPS较SiamRPN++算法有了较大提升,性能较优。改进模块的SiamRAT算法的鲁棒性和FPS有了较大提升,平均精确度达到了0.909。说明SiamRAT算法能够很好地适应于牛场无人机跟踪环境。
To improve the real-time performance and robustness of unmanned aerial vehicle target tracking algorithms in cattle farms,with the unmanned aerial vehicle cattle tracking events as the research object,a target tracking algorithm based on siamese tracker with residual accumulation template(SiamRAT)was proposed in the experiment.The lightweight convolutional network MobileNetV2 and template updating mechanism based on anchor box ratio changes was used to improve the real-time performance of algorithms,and the residual accumulation template with confidence and multi-peak Euclidean distance detection module was used to solve the tracking drift issues caused by similar cattle interference.Finally,the SiamRAT algorithm was compared with SiamRPN++,SiamDW,DaSiamRPN,SiamRPN,and ECO-HC algorithms on a test dataset consisting of cattle videos collected by drones and videos with the same attributes in the VOT2018 dataset.The performance was evaluated using average accuracy,robustness,and frame per second(FPS)as indicators,and analyzed the contribution of improvement modules(including residual accumulation template,high confidence update,and peak distance detection)to the SiamRAT algorithm.The results showed that compared with SiamRPN++,SiamDW,DaSiamRPN,SiamRPN,and ECO-HC algorithms,SiamRAT algorithm had the best robustness,and the average accuracy decreased slightly,but it still ranked the second among all algorithms;FPS had significantly improved compared to SiamRPN++algorithm,with beter performance.The robustness and FPS of the improved SiamRAT algorithm had been significantly improved,with an average accuracy of 0.909.It indicated that the SiamRAT algorithm could be well applied to the tracking environment of unmanned aerial vehicles in cattle farms.
作者
鲁宇
杨颜博
LU Yu;YANG Yanbo(College of Information Engineering,Inner Mongolia University of Science&Technology,Baotou 014000,China)
出处
《黑龙江畜牧兽医》
CAS
北大核心
2024年第2期33-42,共10页
Heilongjiang Animal Science And veterinary Medicine
基金
内蒙古自治区自然科学基金项目(2020LH06006)
内蒙古教育厅基金项目(0406082219)
内蒙古自治区科技厅重大专项(2019ZD025)。
关键词
目标跟踪
模板更新
孪生网络
轻量卷积网络
无人机
跟踪漂移
相似干扰
target tracking
template update
siamese network
lightweight convolution network
drones
tracking drift
similar interference
