Robot Ground Recognition and Prediction Based on Feature Fusion and Bi-LSTM

The ground perception and recognition which have a wide range of application prospects are good for the robot to improve the efficiency of the navigation process. This paper proposed a CNN-Bi-LSTM algorithm based on feature fusion, which effectively combined local features, global features and reverse features in time series data. First, a one-dimensional convolutional layer was used to fuse adjacent local features. After the processing of the max-pooling layer and the Relu function, a bidirectional LSTM was used to extract global time series features and reverse features. Finally, the Softmax function was used for activation, and the classification of nine different road surfaces was realized successfully. The experimental results show that the accuracy of the CNN-Bi- LSTM algorithm is 83.2%, which is higher than that of LSTM, CNN-LSTM, and CNN-GRU-LSTM models.