Abstract:
Automatic detection and recognition of mining roads can not only improve mining transportation safety, but also enhance production efficiency. In order to automatically detect mining roads, an experimental mine is selected for the study, and professional equipment is used to collect road images of the mining area. Subsequently, histogram equalization is used to improve image quality, and median filtering is introduced for data denoising to construct a detection dataset. In the detection model, a bilateral segmentation network model is adopted and improved in three aspects, namely detail branch, semantic branch, and feature recovery. The results show that the maximum accuracy of the detection model is 97.82%, which is 10.63%, 7.07%, 6.17%, and 5.55% higher than the maximum values of the four comparison models, respectively. The maximum
F1 value of this model is 0.993, indicating good overall performance. The maximum Dice coefficient is 0.975, which is closer to 1, and the average mean square error and mean absolute error are 1.204 and 1.110, respectively. In addition, in terms of computational consumption, the maximum values of the model’s time consumption, CPU utilization, and memory occupancy are 83 ms, 11.37%, and 10.95%, respectively, which are significantly better than the comparison algorithms. The mining road detection model designed for research has good performance and can provide technical support for road recognition and guidance for mining trucks.