基于时空网络的变电站巡检机器人视觉导航避障研究OA北大核心CSTPCD

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.