Research On Unmanned Environment Vision Perception Algorithm In Road Scenes

The visual perception algorithm of unmanned environment in road scenes is the main source of information for control and decision making of unmanned systems,and accurate and real-time visual perception algorithm is an important guarantee for safe operation of unmanned systems.However,complex and changing road scenes and external factors such as haze can increase the difficulty of environment perception.To improve the accuracy and real-time of the unmanned environment visual perception algorithm,the following research is conducted in this paper.First,the road scene image dehazing algorithm in hazy weather is investigated.The GCANet dehazing algorithm is preferred,and for its failure to consider local pixel correlation,the adjacency information-assisted prediction module is proposed to enhance the connection between adjacent pixels,and the multi-scale featureconstrained decoding module is proposed to enrich the decoding features and recover clear and high-quality images,which lays the foundation for the subsequent research on object detection and instance segmentation algorithms.Second,the object detection algorithm in visual perception of unmanned environment is improved.The FCOS algorithm is preferred,and the cross-scale improvement strategy and Io U-better strategy are proposed to improve the cross-scale expression capability and object localization accuracy of the algorithm.A feature degradation strategy is proposed to extract more significant features for each type of object,drawing on traditional detection algorithms.And the re-detection strategy is proposed to deepen the object feature understanding.The feature sharing strategy is also proposed to optimize the real-time performance of the algorithm.Finally,two object detection algorithms,Stronger-FCOS and Faster-FCOS,are constructed for different scenarios.Among them,the Faster-FCOS algorithm improves the detection speed by 2.5 FPS and the detection accuracy by 3%,which has obvious accuracy and speed advantages compared with typical object detection algorithms.After that,the Polar Mask instance segmentation algorithm is improved.For its shortcomings in adapting to segmentation tasks in high-resolution road scenes,a fully densely connected classification and ray-distance regression network is proposed,and a correlation-based ray-distance Encoder-Decoder module is proposed to enhance the correlation between object contours and categories and improve the segmentation accuracy.Experimental results show that the improved algorithm proposed in this paper achieves a 2.1% improvement in segmentation accuracy on the Cityscapes dataset,which is slightly better than the Mask R-CNN algorithm and has a significant speed advantage.Finally,a balanced study of multi-task instance segmentation algorithms based on object detection and contour segmentation tasks is conducted.To address the limitations of the current multitask network structure and loss function,a new multitask network MTBNet is proposed.firstly,the shared network structure and each task branch are designed modularly,and secondly,the GHM loss function and multitask loss equalization strategy are introduced to improve the algorithm’s ability of difficult sample mining and dynamically equalize the multitask loss.The experimental results show that MTBNet achieves 29.2% segmentation accuracy and 17.6 FPS segmentation speed on the Cityscapes dataset,which is better than Mask R-CNN and other typical example segmentation methods.In summary,the MTBNet algorithm finally proposed in this paper can better balance accuracy and real-time performance in road scenarios,which is an important reference value for promoting the industrialization of unmanned systems.