Accurate and rapid counting has been one of the most important steps on sea foods in modern aquaculture. Taking the sea foods in a real bottom-sowing aquaculture environment as the research object, this study aims to guarantee the cost-saving, highly efficient and easy counting on the multi-category sea foods using video multi-target tracking. The underwater videos of sea foods were taken as the data source. Firstly, the YOLOv7 algorithm with excellent performance was used to realize the target detector of sea treasure, providing the input for the multi-target tracking. Specifically, the image data was captured under real underwater using the "2017 Underwater Robot Picking Contest, URPC2017", and then used to train the YOLOv7 model. 17,655 images contained the holothurians, echinus, and scallops. The images were randomly divided into the training set of 14055 images in the ratio of 8:1:1, the verification set of 1800 images, and the test set of 1800 images. The training setting of the YOLOv7 detector was set as the image adaptive size of 640*640, the initial learning rate of 0.01, the momentum of 0.9, the weight decay of 0.0005, the batch size of 16, the number of training rounds of 300. One test was performed every 10 rounds of training round. The operating system was selected the Ubuntu18.04 for Haizhenpin detector training, the deep learning framework was PyTorch, the experimental processor was AMD Ryzen Threadripper 1920X 12 cores, and the graphics card was NVIDIA GeForce RTX2080. Secondly, the characteristics of high similarity were combined with the unclear appearance of the same type of sea foods in the real breeding environment. The multi-target tracking of the BYTE algorithm was used for the reference. A multi-category trajectory generation and a counting strategy were then designed for the sea treasure tracking using the trajectory ID number. Finally, an optimal combination of the indicators was also evaluated on the performance of the improved model. The test results show that the average counting precision, mean absolute error (MAE), root mean square error (RMSE), and frame rate were 91.62%, 5.75, 6.38, and 32 frames/s, respectively. All indicators were better than those of the current YOLOv5+DeepSORT, YOLOv7+DeepSORT, YOLOv5+BYTE, and YOLOv7+BYTE. Especially, the average counting accuracy and frame rate index ratio of YOLOv5+DeepSORT were improved by 29.51, and 28 percentage points than before. The MAE and RMSE of the improved model were reduced by 19.50 and 12.08, respectively. The quantity of underwater sea foods was effectively measured in the modern fishery. The finding can provide the technical reference for the production measurement and the scientific decision-making on the intelligent management of aquaculture. In addition, the underwater dataset was used to train the detection model under the same environment, in order to reduce the false detections caused by different conditions. Nevertheless, the underwater environment changes greatly, due to various influencing factors, such as lighting. Statistical counting was then performed to collect the video of marine treasures under the same environment every time during training. Therefore, the current model can be expected to improve on the different underwater environments in the future.

Population census of sea cucumbers can serve as the fundamental and operational metric in contemporary aquaculture. The essential data can also be used for production management, feed administration, and commercial transactions. However, manual sampling and counting cannot fully meet the large-scale production in recent years. Multiple operational challenges still remain in the industry. These conventional approaches can also suffer from three principal limitations: 1) Exorbitant labor requirements can lead to unsustainable operational costs; 2) Suboptimal counting speeds can result in processing bottlenecks during high-volume operations; 3) Inherent human error can cause measurement inaccuracies. The continuous expansion of sea cucumber farming has been exacerbated by the current production scales in the economic and logistical practice. It is the urgent industry demand for automated, intelligent counting for delivering accurate, real-time population data. In this study, an intelligent counting framework was presented using advanced machine vision on unmanned surface vehicles (USVs). The solution framework was integrated into three synergistic technical modules, including object detection, target tracking, and automated counting. Among them, the YOLOv8s-BB model was developed for optimal detection. The underwater object recognition was significantly improved after a key architectural adjustment. Firstly, the BiFormer convolutional attention module (BCAM) was designed to strategically combine a Bi-level routing with the spatial attention mechanism. This hybrid attention module was then integrated at the terminal layer of the YOLOv8s Backbone network. Secondly, the dedicated BiFormer modules were inserted after each C2f block in the Neck architecture. The feature extraction and fusion were enhanced for the underwater environments with variable lighting conditions and turbidity. In the tracking module, an advanced tracking algorithm, TriSORT, was featured by the triple-matching, in order to effectively solve the target limitations of the conventional ByteTrack in aquatic settings. Specifically, the TriSORT algorithm was implemented on a three-stage cascade matching between detection and predicted boxes. The high-confidence detection was introduced to enhance the tracking accuracy and robustness. The counting module was implemented to compare the traditional line-crossing and the inactive trajectory removal. The superior performance was achieved by dynamically filtering the transient detections. Only validated trajectories were maintained to eliminate the common counting artifacts prevalent in aquatic environments. Comprehensive experimental evaluations demonstrated that the superiority of the framework: 1) The YOLOv8s-BB detector was achieved in the state-of-the-art performance with 88.9% mean average precision (mAP), 77.8% recall, and 84.2% F1 score, thus outperforming existing YOLO model variants: YOLOv8s, YOLOv7-tiny, YOLOv9s, and YOLOv11s models; 2) The TriSORT tracking algorithm was enhanced the tracking robustness, with the 74.00% multiple object tracking accuracy (MOTA) and 85.03% IDF1 score, indicating the improvements of 6.55 and 5.54 percentage points, respectively, over ByteTrack; 3) The removal counting of the inactive trajectory was achieved in the superior performance with an average counting accuracy of 95.46% and a mean absolute error of 1.90. The superior performance was achieved in all evaluated metrics over conventional line-crossing counting. The great contribution was also gained in the aquaculture technology. A fully optimal detection-tracking-counting pipeline was established for the precise, efficient, and automated assessment of the sea cucumber population. The finding can provide reliable, data-driven support to the critical aquaculture operations, including biomass estimation, feeding optimization, and harvest planning. The potential applications can also be extended to marine species.