Terrain Recognition Based On Deep Transfer Learning And Semantic Segmentation

In an unknown and unstructured environment,mobile robot needs to perceive the surrounding environment,so that it can carry out path planning in time,so that the robot can walk autonomously and smoothly.At present,most convolutional neural network structures are too complex and the number of model parameters is too large,so they are not suitable for mobile terminal deployment.To solve this problem,this paper introduces MobileNetV3,a lightweight network,to reduce the number of parameters while ensuring the accuracy of the model.Due to the staggered and complex outdoor environment,the environment encountered by mobile robots during walking is not unchanged,and the effect of single terrain recognition algorithm is not ideal.In order to improve the recognition accuracy of complex and unstructured terrain,semantic segmentation of images can be carried out to achieve pixel-level classification.The specific work of this paper is as follows:To solve the problem that there are few datasets available for terrain classification at present,a new dataset GXU-Terrain6 is established.GXU-Terrain6 contains six types of typical terrains identified by the experiment:asphalt,grass,ceramic,floor,mulch and cement.Data is collected and enhanced in different weather conditions.In order to solve the problem of poor accuracy of terrain classification algorithm caused by insufficient training data,and to reduce the number of model parameters to make the model suitable for application to mobile terminal,this paper proposes a terrain recognition algorithm based on deep transfer network MobileNetV3.The large-scale shared parameters of the pre-training model network in the source domain are transfered,and the new training network is initialized according to the weight of the pre-training model,and the prior knowledge on the large dataset is inherited.Then we train and learn on the self-built data set,fine-tune the parameters in the network through the target domain sample data,and then get the required lightweight classification model,which can be deployed to the mobile terminal.The environment captured by mobile robots in the process of walking outdoors is complex and changeable,and the recognition ability of a single terrain recognition algorithm is limited.To solve this problem,this paper proposes an improved lightweight semantic segmentation algorithm of DeepLabv3+ Network.In the improved semantic segmentation algorithm,MobileNetV3 is used to replace the original model backbone network for feature extraction,depthwise separable convolution and attention mechanism block are introduced,and group normalization method is adopted at last.The improved algorithm can reduce the number of model parameters and the running time,while maintaining the performance of the algorithm,and achieve a good segmentation effect on the Cityscapes dataset.