【作者】 孙杰；

【导师】 孙月平；

【作者基本信息】 江苏大学 ， 电子信息（专业学位）， 2024， 硕士

【摘要】 自动投饵船能按照事先规划的路径航行并同步完成投饵作业,可大幅减轻渔民劳动强度、提升投饵质效,提高河蟹养殖效益,但在河蟹养殖期间,蟹塘环境存在复杂多变的情形,由于养殖需要,渔民会在蟹塘中不定期放入竹竿、蟹笼、增氧机等养殖装置,这些装置可能会摆放在事先规划好的投饵船的作业路线上,给投饵船的正常作业带来较大安全隐患,影响投饵船的工作效率。因此,亟需研究投饵船蟹塘障碍物检测定位与自主避障方法,以满足其安全自主巡航作业需求。本文采用深度学习目标检测算法对投饵船前方的障碍物进行实时检测,通过深度相机获取障碍物的空间位置信息,并基于改进人工势场法进行避障路径规划,实现投饵船的自主避障,提高投饵船作业安全性和工作效率。本文的主要研究内容如下:(1)设计了基于改进YOLOv5s的轻量化蟹塘障碍物检测模型。考虑现有模型参数量和计算量较大、模型复杂度高,无法满足投饵船实时检测的需求。改进模型从平衡检测速度和检测精度的角度出发,将Shuffle Net V2网络作为主干网络实现模型的轻量化设计,同时引入SENet注意力机制增强对蟹塘障碍物目标的特征感知;接着引入SPPFCSPC模块,加强不同尺度下蟹塘障碍物的检测效果;最后采用SIo U损失函数加速模型收敛,提高模型的检测准确性。消融实验表明,改进模型的参数量和计算量分别为2.628M和3.3GFLOPs,模型大小为5.5MB,单张图像的检测时间仅15.2ms,检测速度提升了44.5%,平均精度均值为93.3%。与其他主流算法的对比实验表明,改进模型在参数量、计算量、检测速度和检测精度上保持了最佳平衡,更加适合投饵船的应用需求,具有明显的优势。(2)研究了基于深度相机的蟹塘障碍物空间定位方法。在利用改进模型检测到障碍物后,结合Real Sense D435i深度相机获取的深度信息,将像素坐标系中的障碍物定位点映射到蟹塘坐标系下的三维空间定位点,实现蟹塘障碍物的空间定位和宽度测量。竹竿、蟹笼、增氧机3类障碍物的定位精度实验表明,在2～10m范围内测得深度相机与3类障碍物间距离的平均绝对误差最大为0.17m,相对误差为2.39%;测得深度相机与3类障碍物间距离的最大绝对误差为0.36m,相对误差为3.87%。障碍物宽度检测精度实验表明,竹竿、蟹笼和增氧机的宽度测量误差分别集中在0～0.05m、-0.05～0.13m、0.01～0.19m,平均宽度误差分别为0.03m、0.05m、0.10m,验证了该方法具备较高的检测定位精度,为投饵船后续的避障决策提供依据。(3)设计了投饵船的避障路径规划算法。针对传统人工势场法在投饵船避障过程中出现的问题,改进人工势场法首先结合目标检测获取到的障碍物位置和宽度信息,引入自适应目标点解决目标不可达问题;接着针对在避障过程中面临的局部最优解,提出了动态目标点策略用以引导投饵船驶离陷阱区域;最后结合深度相机检测到的障碍物数量提出了自适应调节步长策略,提高避障路径规划的效率。通过MATLAB分别对自适应目标点、动态目标点以及自适应调节步长3种改进策略进行仿真,验证了改进策略的有效性,相较于传统人工势场法,改进人工势场法不仅避免了上述存在的问题,而且避障路径规划时间缩短了26.3%,迭代次数从95次降到78次,避障路径规划的实时性得到进一步提升。在投饵船自主避障效果实验中,改进人工势场法规划的路线在保证投饵船安全的前提下顺利完成了对3类障碍物的自主避障,验证了改进方法的实用性。更多还原

【Abstract】 The demand for Chinese mitten crabs in China is expanding day by day,and the total production of mitten crabs continues to grow rapidly.Industrialization and scaling have become the new trend in mitten crab farming.Due to the worsening aging population in China and the increasing cost of labor,extensive farming methods are no longer able to meet the requirements of modern fisheries.Traditional Chinese mitten crab farming mainly relies on manual feeding by paddling boats,which is labor-intensive,with poor feed distribution and low operational efficiency.Automatic feeding boat can navigate along a pre-planned route and simultaneously perform the feeding operation,significantly reducing the labor intensity for fishermen,improving feeding efficiency,and enhancing the profitability of Chinese mitten crab farming.During the Chinese mitten crab farming process,the environment of the crab ponds is subject to complex and constantly changing conditions.Due to the farming requirements,fishermen may intermittently place breeding equipment such as poles,tarps,and aerators in the crab ponds,these devices may be positioned along the pre-planned route of the feeding boat,posing significant safety hazards and affecting the efficiency of the feeding boat operations.Therefore,it is urgent to research obstacle detection,positioning,and autonomous obstacle avoidance methods for feeding boat in crab ponds to meet their safety requirements for autonomous cruising operations.In this paper,deep learning object detection algorithm was used to detect the obstacles in front of the feeding boat in real time,the spatial position information of obstacles was obtained through the depth camera,and the obstacle avoidance path was planned based on the improved artificial potential field method,so as to improve the safety and work efficiency of the feeding boat.The specific contents are as follows:1)A lightweight crab pond obstacle detection model based on improved YOLOv5 s was designed.Considering that the existing model cannot meet the requirements of real-time detection of feeding boat due to the large number of parameters,calculation amount and complexity of the model,the improved model focuses on balancing detection speed and accuracy.Shuffle Net V2 was adopted as the backbone network for lightweight design.At the same time,Squeeze-and-Excitation attention mechanism was introduced to enhance the feature perception of obstacles in crab pond.Then,the SPPFCSPC module was introduced to strengthen the detection effect of obstacles in crab ponds at different scales.Finally,SIo U loss function was used to accelerate the model convergence and improve the detection accuracy of the model.The ablation experiments showed that the parameters and computation amount of the improved model were 2.628 M and 3.3GFLOPs,respectively,the model size was 5.5MB,the detection time of single image was only 15.2ms,the detection speed was increased by 44.5%,and the average precision was 93.3%.Compared with other mainstream algorithms,the improved model maintains the best balance in terms of parameter count,computational complexity,detection speed and accuracy.It is more suitable for the application requirements of the feeding boat and exhibits significant advantages.2)The spatial location method of obstacles in crab pond based on depth camera was studied.After detecting obstacles using the improved model,combining the depth information obtained from the Real Sense D435 i depth camera,the obstacle location points in the pixel coordinate system were mapped to three-dimensional spatial location points in the crab pond coordinate system to realize the spatial positioning and width measurement of the crab pond obstacles.The positioning accuracy experiments for three types of obstacles,including poles,traps,and aerators,showed that within the 2～10m range,the average absolute error of the distance between the depth camera and the three types of obstacles was 0.17 m,with a relative error percentage of 2.39%.The maximum absolute error of the distance between the depth camera and the three types of obstacles was 0.36 m,with a relative error percentage of 3.87%.The width measurement errors for poles,traps,and aerators were concentrated within the ranges of 0 ～ 0.05 m,-0.05 ～ 0.13 m,and 0.01 ～0.19 m,with average errors of 0.03 m,0.05 m,and 0.10 m,respectively.Above experiments validate that the method has high detection and positioning accuracy,providing a basis for obstacle avoidance decision-making for the feeding boat.3)An obstacle avoidance path planning algorithm for feeding boat was designed to address the issues encountered by the traditional artificial potential field method during the obstacle avoidance process.In the improved artificial potential field method,the obstacle position and width information obtained through obstacle detection model were combined,and adaptive target point was introduced to solve the problem of goal nonreachable.Subsequently,to address the local optimal solution faced during the obstacle avoidance process,a dynamic target point strategy was proposed to guide the feeding boat away from trap areas.Finally,in conjunction with the obstacle quantity obtained by the depth camera,an adaptive adjustment step strategy was proposed to enhance the efficiency of obstacle avoidance path planning.Three improvement strategies,namely adaptive target point,dynamic target point,and adaptive adjustment step size,were simulated with MATLAB to verify the effectiveness of the improvement strategies.Compared to the traditional artificial potential field method,not only the issues existing in the original algorithm was avoided but also the path planning time was reduced by 26.3% in the improved artificial potential field method.The number of iterations was decreased from 95 to 78,leading to a further enhancement in real-time obstacle avoidance path planning.In the obstacle avoidance experiments for feeding boat,the autonomous obstacle avoidance was performed successfully with the path planned by the improved artificial potential field method while ensuring the safety of the feeding boat.The practicability of the improved method is verified.更多还原