面向AUV自主水下对接的视觉检测算法

Visual Detection Algorithm for AUV in Autonomous Underwater Docking

摘要:

针对自主式水下机器人水下对接过程中近距离视觉导引存在的定位精度低、计算时间长的问题。 基于 Cascade RCNN 神经网络模型设计了一种高精度视觉检测算法。首先，基于热力图导向的泊松复制建立混合数据增强策略，在实例级和图像级上扩充训练样本；其次，采用完备交并比损失函数提高算法模型输出边界框的回归准确率；然后，使用学习率余弦退火策略，通过周期性的热重启和衰减使算法模型能概率性地跳出局部最优解并最终趋于全局最优，以进一步提高目标检测精度；最后，使用混合精度策略减少算法模型计算复杂度，提高实时性。水下对接实拍图像目标检测试验表明：该算法模型的平均精度达到 94%，检测速度超过 20 帧/秒。

Abstract:

Aiming at the problem of low position accuracy and long calculation time in near-distance visual guidance of autonomous underwater vehicle during underwater docking，a high-precision visual detection algorithm based on Cascade RCNN neural network model is designed. Firstly，a hybrid data augmentation strategy is established based on heat-map guided poisson copy，so training samples are expanded at both instance-level and image-level. Secondly，the complete intersection of union loss function is used to improve the regression accuracy of the output boundary frame of the model. Then，the learning rate cosine annealing strategy is used to improve the final precision of the target detection by periodic thermal recovery strategy and the global attenuation strategy. Finally， mixed precision strategy is used to reduce the computation complexity of the model and to improve the real-time performance. Experiments on underwater docking images show that the average accuracy of the proposed algorithm is over 94% and the detection speed is over 20 frames per second.