Tuna catch real-time detection by fusing channel pruning with ByteTrack lightweight network
-
摘要: 自动准确收集渔业捕捞数据是电子观察员系统的重要组成部分,然而,由于工作环境的复杂性和跟踪的不稳定性,金枪鱼延绳钓渔获数量自动估计在实践部署中仍存在挑战。本研究设计了一个轻量级计数网络对渔船上的实时视频数据进行自动处理,实现对金枪鱼渔获物的实时跟踪和计数。本研究选择YOLOv5s作为基准网络,首先采用通道剪枝算法对YOLOv5s的主干网络进行修剪,结果表明,剪枝后的模型检测精度mAP0.5~0.95达到68.8%,CPU下检测速度为16.5帧/s（FPS）,与原始模型相比,检测效果基本不变,模型的参数量、模型大小和计算量分别减少了67.2%、66.4%和42.5%,检测速度提高了33.1%。其次,利用ByteTrack算法实现了多目标的实时跟踪,优化了计数区域形状,解决了被跟踪金枪鱼身份（ID）跳变导致的计数偏差问题,10个视频的测试结果表明,该方法的平均计数准确率为80%,视频处理速度为50.7帧/s,满足工业级实时检测要求。综上,该模型具有轻量化、高精度、实时性等优点,可在复杂的工作环境下完成对延绳钓捕捞结果的实时监控,为实现渔业自动化提供思路。Abstract: Automatic and accurate collection of fishery catch data is an important part of the electronic observer system. However, automated tuna longline catch estimation remains challenging to deploy in practice due to the complexity of the working environment and the instability of tracking. In this study, a lightweight counting network was designed to automate the processing of real-time video data from fishing vessels in order to enable real-time tracking and counting of tuna catches on board fishing vessels. YOLOv5s was selected as the benchmark network in this study. The channel pruning algorithm was first used to prune the backbone network of YOLOv5s, and the results showed that the detection accuracy of the pruned model mAP0.5~0.95 reached 68.8%, and the detection speed was 16.5 frames per second （FPS） under CPU, which was basically unchanged compared with the original model. The number of parameters, model size and computation of the model were reduced by 67.2%, 66.4% and 42.5% respectively, and the detection speed was increased by 33.1%. Secondly, the ByteTrack algorithm was used to achieve real-time tracking of multiple targets, optimize the shape of the counting area and solve the problem of counting deviation caused by the jump in the identity （ID） of the tuna being tracked. The test results of 10 videos showed that the average counting accuracy of the method was 80% and the video processing speed was 50.7 FPS, which meets the requirements of industrial-grade real-time detection. In summary, the model had the advantages of light weight, high accuracy and real-time, which could complete the real-time monitoring of longline catch in complex working environments and provide a solution to realizing fisheries automation.
-
Key words:
- tuna individuals /
- channel pruning /
- ByteTrack /
- real-time detection /
- YOLOv5 /
- multi-target tracking
-
-
GOLDEN C D, ALLISON E H, CHEUNG W W L, et al. Nutrition: Fall in fish catch threatens human health[J]. Nature, 2016, 534（7607）: 317-320.
YANG X T, ZHANG S, LIU J T, et al. Deep learning for smart fish farming: applications, opportunities and challenges[J]. Reviews in Aquaculture, 2021, 13（1）: 66-90.
VAN HELMOND A T M, MORTENSEN L O, PLET-HANSEN K S, et al. Electronic monitoring in fisheries: lessons from global experiences and future opportunities[J]. Fish and Fisheries, 2020, 21（1）: 162-189.
BICKNELL A W, GODLEY B J, SHEEHAN E V, et al. Camera technology for monitoring marine biodiversity and human impact[J]. Frontiers in Ecology and the Environment, 2016, 14（8）: 424-432.
BROWN C J, DESBIENS A, CAMPBELL M D, et al. Electronic monitoring for improved accountability in western Pacific tuna longline fisheries[J]. Marine Policy, 2021, 132: 104664.
PROBST W N. How emerging data technologies can increase trust and transparency in fisheries[J]. ICES Journal of Marine Science, 2020, 77（4）: 1286-1294.
LIU Y Q, CHU H Y, SONG L M, et al. An improved tuna-YOLO model based on YOLOv3 for real-time tuna detection considering lightweight deployment[J]. Journal of Marine Science and Engineering, 2023, 11（3）: 542.
CAI K W, MIAO X Y, WANG W, et al. A modified YOLOv3 model for fish detection based on MobileNetv1 as backbone[J]. Aquacultural Engineering, 2020, 91: 102117.
IQBAL M A, WANG Z J, ALI Z A, et al. Automatic fish species classification using deep convolutional neural networks[J]. Wireless Personal Communications, 2021, 116（2）: 1043-1053.
BRADLEY D, MERRIFIELD M, MILLER K M, et al. Opportunities to improve fisheries management through innovative technology and advanced data systems[J]. Fish and Fisheries, 2019, 20（3）: 564-583.
TSENG C H, KUO Y F. Detecting and counting harvested fish and identifying fish types in electronic monitoring system videos using deep convolutional neural networks[J]. ICES Journal of Marine Science, 2020, 77（4）: 1367-1378.
ZHANG S, YANG X T, WANG Y Z, et al. Automatic fish population counting by machine vision and a hybrid deep neural network model[J]. Animals, 2020, 10（2）: 364.
MICHELINM, ELLIOTTM, BUCHERM, et al. Catalyzing the growth of electronic monitoring in fisheries[R]. California Environmental Associates, 2018.
WANG D D, HE D J. Channel pruned YOLO V5s-based deep learning approach for rapid and accurate apple fruitlet detection before fruit thinning[J]. Biosystems Engineering, 2021, 210: 271-281.
JIANG P Y, ERGU D, LIU F Y, et al. A review of Yolo algorithm developments[J]. Procedia Computer Science, 2022, 199: 1066-1073.
MAHASIN M, DEWI I A. Comparison of CSPDarkNet53, CSPResNeXt-50, and EfficientNet-B0 Backbones on YOLO V4 as Object Detector[J]. International Journal of Engineering, Science and Information Technology, 2022, 2（3）: 64-72.
ZHANGY F, SUNP Z, JIANGY, et al. ByteTrack: Multi-object tracking by associating every detection box[C]//17th European Conference on Computer Vision. Tel Aviv, Israel: Springer, 2022: 1-21.
WELCH G, BISHOP G. An introduction to the Kalman filter[M]. Chapel Hill, NC, United States: University of North Carolina, 1995.
KUHN H W. The Hungarian method for the assignment problem[J]. Naval Research Logistics Quarterly, 1955, 2（1/2）: 83-97.
计量
- 文章访问数: 662
- PDF下载数: 62
- 施引文献: 0
下载:
