Object detection algorithm based on improved YOLOv8 for drill pipe on coal mines.

Gas extraction is an important measure for coal mine gas disaster control. Its effect is closely correlated to the drilling depth. The existing methods usually determine the drilling depth by manually counting the number of drill pipes, and the number of drill pipes can be automatically counted by object detection and real-time tracking algorithms. An improved object detection model was proposed for the problem of the poor performance of the object detection algorithm due to such interference factors as bright light, low illuminance and heavy dust and mist in coal mines. In terms of data augmentation, the ACE dehazing algorithm is introduced to improve image quality. In order to solve the problem of leak detection caused by the irregular shape that appears due to the interference of bright light, the deformable convolution DCNv2 module was integrated in the C2f module to make the sampling points of the convolution kernel diffuse irregularly, so as to fully extract the shape features of the drill pipe and then improve the detection rate of the model. For the problem of too low confidence of the model in detecting drill pipes due to uneven illumination, the attention paid by the model to the features of the drill pipe could be improved by embedding the SimAM non-parametric attention mechanism module in the backbone network, which can further improve the confidence of the drill pipe. For the problem of low average category detection accuracy caused by the changeable environment of the underground drilling site, the dynamic head was used to improve the ability of the model to extract the features of the drill pipe in scale, space, and channel, and improve the average category detection accuracy of the drill pipe. To address the issue of diverse angle differences between predicted and real boxes, CIoU loss function is replaced with the SIoU loss function. Finally, the improved detection algorithm was verified with the homemade drill pipe dataset. The experimental results showed that: the improved model effectively alleviated the problem of partial leak detection of the original network for scenes such as heavy dust and mist and uneven illumination; the recall rate increased by 4.9%; the mean average precision was improved by 5.3%. At the same time, it maintains a high real-time performance (the FPS is 117), providing the basis of the drill pipe detection model for the application of real-time tracking of the number of drill pipes.