Road negative obstacle detection in open⁃pit mines based on multi scale feature fusion

The existence of negative obstacles such as potholes and road cave⁃ins on unstructured roads in open pit mines can easily lead to vehicle rollover or trapping，and the development of autonomous mining truck driving in re⁃ cent years has made negative obstacle detection crucial. In this study，the characteristics of road negative obstacles in open⁃pit mining area are comprehensively analyzed，and a lightweight object detection model based on machine vision is constructed. Firstly，the negative obstacle data set of open⁃pit mining area is constructed by field sampling and input into the target detection model. Secondly，the image is normalized and activated by mobileNetv3 network. After obtai⁃ ning the output features，the continuous up⁃sampling and feature pyramid stacking are performed to complete multi⁃ scale feature extraction. Finally，the multi⁃scale features are classified and the bounding box regression is performed to achieve negative obstacle detection. The model introduces a depth⁃separable convolution method in the feature pyramid module to reduce the computational effort of network feature extraction and fusion. By adjusting the loss function and learning rate dynamically and optimally，the accuracy of negative obstacle target detection is improved. In the post⁃ processing stage of negative obstacle detection，a non⁃maximum suppression optimization algorithm is proposed in the post⁃processing stage of negative obstacle detection to solve the problem that the negative obstacles are obscured and the detection frame is not localized with high accuracy. The experimental results show that the research meth⁃ od can effectively identify negative obstacles on unstructured roads in open⁃pit mining areas in complex backgrounds. The mine road negative barrier detection precision，recall and mAP reach98. 86%，89. 58% and 92. 59% with a real time detection speed of 47.3 fps. It also has a better overall performance than that of mainstream target detec⁃ tion networks such as Yolov3，RetianNet，SSD，Faster RCNN etc. A quantitative analysis of the non⁃maximal suppres⁃ sion optimization algorithm has also shown that the improved algorithm is effective in improving the positioning accura⁃ cy of the detection boxes compared to the traditional algorithm and is also very adaptable.