An accurate and rapid detection is a high demand to identify the river crab and bait in underwater complex environments. In this study, an improved YOLOv5s was proposed to detect the river crab and bait for the high precision, speed and simple modeling. GhostNet and GhostBottleneck structure were used to rapidly extract the network features of the model. The inference and detection speed were improved with the less complexity and amount of model computation. The model was deployed on the Android, in order to meet the needs of mobile application scenarios for river crab detection. The BiFPN structure was applied in the neck network of YOLOv5s. The fusion ability of the model was enhanced for different scale targets, the robustness of the model, and especially the detection performance of the network for small targets. A lightweight attention mechanism was used in the backbone feature extraction of the YOLOv5s network. The CA attention mechanism module was employed to improve the detection probability of the target region. The relevant feature information was matched with the target channel. The invalid information was suppressed to strengthen the network attention to river crab and bait. The lightweight YOLOv5s-GhostNet model was compared with the YOLOv5s, ShuffleNetv2-YOLOv5s, and MobileNetV3-YOLOv5s models. The results showed that the target average precision of YOLOv5s-GhostNet model was reduced by only 4.6 percentage points, compared with YOLOv5s. The lightweight network was effectively maintained the detection precision of the model for the higher detection speed of river crabs and bait. Ablation experiments were carried out to verify the backbone feature extraction network, feature fusion network BiFPN, and CA attention mechanism. The computational amount was reduced by 48.7%, whereas, the detection speed increased by 27.2%, after improving the backbone feature extraction network with GhostNet. The model complexity was reduced to elevate the detection speed. The BiFPN structure was used to compensate for the loss of mAP that caused by the lightweight network, particularly for the detection rate of river crabs and bait. The addition of CA attention mechanism was improved the anti-interference and feature extraction, while the mAP also increased. River crab and bait datasets were established using laboratory and actual crab pond environments. The improved model was achieved in an average precision of 96.9% and a computational volume of 8.5 GFLOPs, which was higher in the detection precision and smaller in computational volume, compared with the current mainstream single-level target detection algorithms for anchor boxes (e.g., SSD and YOLOv3). The average precision of the model was improved by 2.2 percentage points, while the computation and model memory were reduced by more than 40%, compared with YOLOv5s. The model before and after improvement was deployed to Android phone for testing. It was found that the improved model shared the average detection speed of 148ms/frame on Android phone, which increased by 20.9%, and maintained the better detection, compared with the original. The detection rates of river crabs and bait in the improved model increased by 4.6 and 5.8 percentage points, respectively, which was effectively improved the detection rates than before. Therefore, the improved model can be expected to balance the performance requirements of Android phone for model detection precision and speed. The finding can provide the guidance for the precise determination of baiting amount in river crab culture.

Obstacles have posed a great challenge to the unmanned working vessel in the crab pond. Once the obstacles appear on the route of the unmanned working vessel, it is easy to collide with the vessel. It is very necessary to quickly identify and locate the obstacles in the crab pond, in order to improve the efficiency and safety of the unmanned working vessel. In this study, an improved YOLOv5s model was proposed to detect the obstacles in the crab pond. The obstacles were then located to combine with the depth camera. Firstly, the lightweight network ShuffleNetV2 was used as the backbone feature extraction network, and the depth-separable convolution and channel mixing strategies were adopted to greatly reduce the model volume, in order to accelerate the detection speed for high accuracy. Secondly, the SE attention mechanism was introduced to enhance the feature perception of obstacles in the crab pond, particularly without increasing the amount of computation. Thirdly, the SPPF module was improved into the SPPFCSPC module. The target features were extracted under different receptive fields, in order to enhance the detection of obstacles in crab ponds at different scales. Finally, the SIoU loss function was adopted to further accelerate the model convergence for high accuracy. According to the color images obtained by the RealSense D435i depth camera, the obstacle was detected to locate the pixel coordinates of the center point of the obstacle in the coordinate system of the crab pond using coordinate conversion. The width of the obstacle was obtained simultaneously. The experimental results showed that the better recognition of the improved model was achieved in the pole, tarp, and aerator, where the confusion matrix output was observed by the model on the test set. The number of parameters and calculation amount of the improved YOLOv5s model were reduced by about 62.8% and 80.0% respectively, compared with the original. The model size was only 5.5 MB, and the volume was reduced by about 61.8%. The detection speed increased by 44.5%, and the inference speed of a single image reached 15.2 ms, where the mean average precision reached 93.3% for obstacles in the crab pond. Compared with the YOLOv5s-MobileNetV2, YOLOv5s-GhostNet, YOLOv7 and YOLOv8, the improved model had maintained the best balance in terms of parameter number, computation amount, detection speed and detection accuracy. A series of positioning and width measurement tests were conducted to verify the location accuracy of the improved model. Three typical obstacles were located in a crab pond at the fishery Science and Technology Demonstration base in Jintan District, Changzhou, Jiangsu Province. The test results showed that the average absolute error and average relative error of the distance between the three types of obstacles and the camera were 0.16 m and 2.26% in the range of 2-10 m, respectively, and the maximum absolute error and maximum relative error were 0.35 m and 3.74%, respectively. The width measurement errors of the pole, crab trap and aerator were concentrated in 0-0.05 m,-0.05-0.13 m and 0.01-0.19 m, respectively. The average errors were 0.03, 0.05 and 0.10 m, respectively, and the average relative errors were 34.1%, 7.5% and 5.0%, respectively. Finally, the depth camera was fixed in the front of the unmanned working vessel, in order to detect the three types of obstacles in the crab pond several times. The detection and positioning effects of the model were verified in the actual crab pond environment. The improved model was used to successfully detect the obstacles, in order to export the category, confidence, three-dimensional coordinate and width information. In summary, the improved model can fully meet the requirements to detect and locate the common obstacles in the crab pond environment. The finding can provide an important reference for the autonomous obstacle avoidance and cruise operation of the unmanned working vessel.

Most of the crabs are fed manually at present in China. However, manual feeding cannot fully meet the industrial requirements of mass production, particularly with an increasingly aging labor force. Fortunately, automatic navigation can be expected to promote the development of unmanned feeding boats. This study aims to realize the real-time path planning of an unmanned feeding boat in a crab pond environment with static and dynamic obstacles. A dynamic path planning was proposed for the unmanned feeding boat using the fusion of the improved A* and the artificial potential field (FIA*-APF). Better performance was achieved to solve the excessive turning points, large path curvature, and numerous redundant points in the traditional A* path planning. Firstly, the evaluation function was optimized to find the end of the path. Then, the turn penalty function was introduced to remove the redundant points for the smaller number of turning points. The path was then smoothed using a B-spline curve. Finally, the gravitational and repulsive functions of the artificial potential field were improved to obtain the global paths using the improved A*. The gravitational paths were used to treat the local optimization and unreachability in the artificial potential field. Two aquaculture pond environments with static and dynamic obstacles were established to evaluate the performance of the traditional artificial potential field (TAPF) , the fusion with the A* and artificial potential field (TA*-APF), as well as the FIA*-APF. Planning speed, number of nodes and path length were selected as the evaluating indicators. Each experiment was carried out 20 times in every single environment. The simulation results showed that the average planning time in the two environments was 17.23% and 51.96%, respectively, compared with the TAPF and the TA*-APFs. Furthermore, the average node numbers were 50.64% and 65.03% less than that, respectively. The average path length was 2.82% less than the TA*-APF. The results of 10 crab pond tests showed that the planning time was only 38.16% and 62.42% of the TAPF and the TA*-APFs, respectively. The average path lengths were 29.13% and 10.15% less than that, respectively. In addition, the TAPF and TA*-APFs planned paths with corners greater than 60° were 3.28 and 2.62 times greater than FIA*-APF, while the corners greater than 100° were 3.73 and 1.67 times greater than FIA*-APF, respectively. The path planned was safer and smoother after optimization. The finding can provide a sound reference for the navigation development in unmanned feeding boats.

At present, the cleaning of water plants in crab pond is primarily relied on manual operations, which is characterized by high labor intensity and low efficiency. The automatic water plant cleaning boat can greatly reduce the labor intensity and improve the operation efficiency, and determining the water plant distribution in crab pond is the key basis for planning the efficient operation path of the cleaning boat. Addressing the challenges posed by the high similarity between water plants and shoreline vegetation, as observed in drone-captured images of the crab pond, which complicates the accurate distinction between them, a region segmentation and localization method for water plants in crab pond based on improved YOLOv8n-seg was proposed in this paper. The focus of the improved model lies in the reduction of the model size and the enhancement of the recall rate. Firstly, in order to diminish the model size, the lightweight HGNetv2 (hierarchical graph network) based on RT-DETR (real-time detection transformer) was used as the backbone feature extraction network. Secondly, the neck network was reconstructed based on the lightweight structure of Efficient Rep to reduce the number of parameters and enhance the multi-scale feature fusion ability of the mode. Finally, the SegNext attention mechanism was introduced in the feature extraction layer to enhance the sensitivity of the model to the water plant area. In order to eliminate the redundant regions generated by the model in the recognition process and further improve the segmentation accuracy, the binarization processing was used to optimize the segmentation results, and the image processing algorithm was combined to screen the area of the water plant area. After coordinate conversion, the precise longitude and latitude coordinates of water plant contour were obtained. The experimental results showed that the improved model exhibited a strong discriminatory and segmentation effect on water plants in crab ponds. The parameter number, calculation quantity and model size of the improved model were only 1.49 M, 8.4 GFLOPs, and 3.27 MB, respectively. Compared with the original YOLOv8n-seg model, the parameter number, calculation quantity and model size of the improved model were reduced by 54.3%, 30.6% and 49.7%, respectively. The model before and after improvement was deployed to the Jetson Orin Nano embedded AI computer for testing. It was found that the preprocessing speed of the improved model was almost the same as that of the YOLOv8n-seg model. The inference speed, post processing speed and segment speed were increased by 30.6%, 46.4% and 32.3%, respectively. In addition, the recall, precision, and mean average precision of the improved model achieved values of 91.5%, 89.3% and 95.6%, respectively. The improved model achieved the best balance of calculation, parameters, recall and detection accuracy comparing with YOLOv5s-seg, YOLOv8s-SwinTransformer, YOLOv8s-seg and YOLOv8n-GoldYOLO models. The coordinate conversion test showed that the minimum distance error of water plant location accuracy was 0.13m, the maximum distance error was 0.33 m, and the average distance error was 0.22 m, which verified that the improved model can meet the requirements of water plant region segmentation and localization in crab pond. The finding can provide an important reference for the automatic operation path planning of water plant cleaning boat.

Photovoltaic panel arrays can often shade over the fishing and light complementary ponds. BeiDou/GPS positioning signals are then affected to significantly reduce the autonomous navigation accuracy of unmanned workboats. Additionally, traditional machine vision can easily cause the suboptimal visual navigation of line detection, due to the low robustness. The resulting images are also confined to the variations in the light and shadow, the distribution of aquatic plants, and surface obstacles. In this study, an improved YOLOv8n model was proposed to extract the navigation center line in the fish and light complementary ponds. Firstly, the HGNetV2 network was used as the backbone network, in order to improve the real-time detection. Group normalization (GN) and shared convolution structure were also used to design a lightweight detection head network, in order to reduce the size of the model. Then the SPPF_LSKA module was used as the feature fusion layer to improve the multi-scale feature fusion of the model. Finally, the Wise-IoU (weighted interpolation of sequential evidence for intersection over union) loss function was used to improve the bounding box regression performance and the detection accuracy of remote small targets. The detection frame coordinates of improved YOLOv8n were used to extract the reference points for the positioning of the cement columns on both sides. The lines of the cement columns on both sides were fitted by the least square method. The middle line of the navigation was then extracted by the angle bisection line. Ablation test results showed that the calculation amount, parameter number and model volume of the improved YOLOv8n model decreased by 36.0%, 36.8% and 32.8%, respectively, compared with the original, where the mean average precision (mAP) was 97.9%. The detection speed increased by 42.9%, where the accuracy was 93.1%, and the detection time of a single image was 6.8 ms. Comparison test showed that the improved YOLOv8n model exhibited the smallest size and the highest degree of lightweight, while maintaining a high level of detection accuracy, compared with the YOLOv5s, YOLOv6, YOLOv7, and YOLOv8. Excellent performance was also achieved in detecting the concrete columns. In the positioning test of the navigation center line, the average linear errors between the reference point of the extraction cement column and the manual observation mark were 3.69 and 4.57 cm in the range of 0-5 m and 5-10 m, respectively. The average linear error between the extracted and actual navigation center line was 3.26 cm, with an accuracy of 92%. The real-time test showed that the average extraction speed of the navigation midline was improved by 38.04% (117.01 frame/s) on the Windows11 test platform, compared with the original. Furthermore, the average extraction speed of navigation midline increased by 36.38% (22.34 frame/s) on the Jetson test platform. Consequently, the improved model can fully meet the navigation requirements of unmanned fishing and light complementary pond boat. The findings can provide a strong reference for the subsequent research on the visual navigation system of operation boats.