Drivable Area Detection In The Scene Of Open-pit Mine

The mining industry is the basic industry of economic and social development,providing material support for meeting people’s increasingly better life needs.In order to improve production efficiency and reduce operating costs,open-pit mining companies have developed towards smart mines.The automatic driving mine truck is an important part of the smart mines system.The automatic driving mine truck is also one of the popular directions for the implementation of the automatic driving technology.Environmental perception is a fundamental function of the autonomous driving mining truck,which can provide the important information for the intelligent decision-making and planning control functions.The detection of drivable areas is a key task of the environmental perception of autonomous driving mining trucks,which is crucial for the safe driving of autonomous driving mining trucks in the scene of open-pit mine.In the scene of open-pit mine,the drivable area has no obvious boundary and is irregular in shape,with puddles and dust.These characteristics of the drivable area bring difficulties to the detection task of the drivable area.At present,in the scene of open-pit mine,there are still some problems such as incomplete detection of the interior of the drivable area,inaccurate detection of the boundary of the drivable area,and incomplete or even missed detection of puddles.Therefore,this research topic takes the detection of drivable areas in the scene of open-pit mine as the research object,and the accuracy and stability of the open-pit mine drivable area detection algorithm can be improved under the condition of real-time performance.The main contents of this research topic are as follows:(1)Design the drivable area detection network of the open-pit mine.According to the characteristics and problems of drivable area detection in the s cene of open-pit mine,based on the Fast-SCNN,the open-pit mine drivable area detection network is designed,and its sub-modules are also designed.(2)Improve the feature fusion module.The attention mechanism is added to the original feature fusion module,so that the network pays more attention to learning the spatial position information of the drivable area,which improves the completeness of the interior detection of the drivable area in the scene of open-pit mine.(3)Design the edge information fusion module.The edge information fusion module can fuse the edge features of the drivable area with high-level semantic information,which improves the accuracy of drivable area edge detection in the scene of open-pit mine.(4)Design the loss function.According to the characteristics of class imbalance and image pixel class imbalance in the drivable area detection dataset of open-pit mines,a weighted Dice loss function is designed.It makes the network pay more attention to the learning of puddle features,which improves the accuracy of puddle detection in the scene of open-pit mine.(5)Ablation experiment and comparison experiment.Three groups of ablation experiment are carried out,including the weighted Dice loss function,the improved feature fusion module and the edge information fusion module.The results of the three groups of ablation experiment proved the effectiveness of the improvement measures.In addition,the comparison experiment among the drivable area detection network of the open-pit mine,Mobile Net V3-small and Bise Net V2 are carried out.The comparison experiments show that the drivable area detection network of the open-pit mines can provide more accurate and stable detection results.And the processing speed of the drivable area detection network of the open-pit mines reaches 42.64 FPS,which meets the real-time requirements.