基于时空网络的变电站巡检机器人视觉导航避障研究

Research on Navigation and Obstacle Avoidance of Substation Inspection Robot Based on Spatiotemporal Networks

为提升变电站巡检机器人视觉导航避障效果,将卷积神经网络与循环神经网络相结合,提出一种轻量级时空网络结构。网络以机器人道路场景视频帧为输入,分别从空间域和时域2个维度对可行道路区域特征进行分析挖掘。针对空间域特征,网络先采用多种图像增强技术来丰富道路特征信息,再利用高效率卷积单元、残差连接以及注意力机制等模块构建卷积网络结构,由浅到深提取可行道路空间位置特征。对于时域特征,网络在空间特征基础上引入基于卷积计算的长短期记忆循环网络,保障时序特征提取的同时避免空间特征被破坏,并结合自注意力结构提升网络对关键信息的聚焦度,降低噪声数据干扰。根据所提时空域特征,设计分类回归预测结构,分别对机器人行驶方向及对应偏转角度进行预测,提升机器人导航避障效果。最后,考虑到实际变电站机器人巡检场景的高度相似性,设计特征差分模块来减少对冗余特征重复计算,保障实际应用效率。多个数据集上的实验结果表明:所提方法可以有效提取道路场景信息的时空特征并准确预测机器人下一步动作;与同类型方法相比,该方法具有更高的鲁棒性,可以更高效智能地实现导航避障。

In order to improve the visual navigation and obstacle avoidance performance of substation inspection robots,a lightweight spatiotemporal network structure is proposed by combining the convolutional neural network(CNN)with recurrent neural network(RNN).The network analyzes and explores the features of feasible road areas from both spatial and temporal dimensions based on the robot road scene video frames.For the spatial domain features,the network firstly uses various image enhancement techniques to enrich the road feature information,and then constructs a convolutional network structure using modules such as high-efficiency convolutional units,residual connections and attention mechanisms to extract feasible road spatial position features from shallow to deep.For the temporal domain features,a convolutional computation based long and short-term memory recurrent network is introduced on the basis of spatial features to ensure temporal feature extraction while avoiding spatial feature destruction.At the same time,the network combines self-attention structure to enhance the network’s focus on key information and reduce noise data interference.According to the proposed spatiotemporal features,a classification regression prediction structure is designed to predict the robot’s driving direction and corresponding deviation angle,thus improving the robot’s navigation and obstacle avoidance effect.Finally,considering the high similarity of actual substation robot inspection scenes,a feature differentiation module is introduced to reduce redundant feature repeated calculations and improve practical application efficiency.Experimental results on multiple datasets show that the proposed method can effectively extract the spatiotemporal features of road scene information and accurately predict the robot’s next action.Moreover,compared with similar methods,this method has higher robustness and can achieve more efficient and intelligent navigation and obstacle avoidance.