Research On Free Space Detection And Prediction Technology For UAVs And UGVs

The unmanned platform consisting of a ground unmanned platform and an aerial unmanned platform has the advantages of large sensing range and strong maneuverability,and is widely used in civil and military fields.The understanding of the scene for the unmanned platform is one of the research hotspots in the field of heterogeneous robot cooperation,especially the scene understanding technology in the field environment faces many challenges.In this paper,the high-precision and fast free space detection and prediction technology are studied for the understanding of the cooperative scene of the unmanned platform,and the test is carried out on the unmanned platform composed of Pioneer 3-AT robot and Jingwei M100.The research content of this paper is as follows:(1)A ground local high-precision free space detection algorithm is designed and implemented for the complex environment faced by the ground unmanned platform.Firstly,based on the Inception Residual module,a multi-scale Inception Residual module is proposed to improve multi-scale feature extraction.In addition,through the multi-task learning module,the object detection task and the semantic segmentation task are jointly learned to improve the accuracy of the two associated tasks.Finally,the ground local high-precision free space detection model is constructed and tested,and the validity of the verification model is tested on the real unmanned platform dataset.(2)Aiming at the problem of weak computing power in the airborne unmanned platform,an airborne global fast and free space detection algorithm is designed and implemented.First,the model size and the amount of calculation are reduced by improving the residual non-bottleneck module.Secondly,the module is refined by the asymmetric decoder,which alleviates the loss of precision caused by reduction of the free space decoder size.Finally,the airborne global fast and feasible area detection model is deployed and the validity of the verification model is tested on the real open space unmanned platform dataset.(3)Aiming at the problem of fuzzy prediction of the free space of the unmanned platform,a free space prediction algorithm based on improved ConvLSTM is designed and implemented.Through the free space prediction encoder decoder module and the improved bidirectional Res ConvLSTM U-Net module,a deeper network is constructed,and the time-space representation ability of ConvLSTM is enhanced,thereby improving the prediction accuracy.