Research On Underground Track Recognition Algorithm Based On Improved BiSeNet And Knowledge Distillation

The existing auxiliary transportation system of underground coal mine in China mainly adopts the transportation mode of rail electric locomotive.Due to the complex environment of underground coal mine and improper operation of personnel,underground transportation often causes casualties and property losses.In order to improve this backward situation,it is necessary to further improve the intelligent level of the equipment of underground coal mine transportation system,and the track detection technology in front of the train is an important technology to realize the unmanned driving of rail electric locomotive.The traditional track detection technology has a good detection effect for the linear track with bright environment and ideal shape,but it has limitations in the complex tunnel environment,and the track detection algorithm based on deep learning still has many problems in the detection accuracy and detection rate.Aiming at the above problems,this paper carried out research on mine track detection based on semantic segmentation and knowledge distillation theory,and proposed the corresponding track detection algorithm.The main research contents are as follows:(1)In view of the current underground track detection model does not fully consider the characteristics of track strip and its deficiency in detection accuracy,an improved BiSeNet real-time semantic segmentation network for underground coal mine track detection is proposed.Firstly,the improved overall structure of the segmentation network is presented.Secondly,according to the characteristics of the track,the sub-network feature fusion module is put forward to aggregate the features of different depths and further refine the features.Finally,the underground track training data set is constructed to carry out experimental verification.The results show that the proposed method achieves an average intersection ratio of 72.8% in the collected data set of underground track environment,which meets the requirement of semantic segmentation accuracy.(2)In order to solve the problem that the detection rate of the track detection model is slow and difficult to meet the requirements of application in the actual production environment,an algorithm of underground track detection based on knowledge distillation is proposed.The algorithm uses knowledge distillation to compress the model.The ESPNet network with good real-time performance in semantic segmentation network is selected as the student network,and the underground coal mine track detection network based on improved BiSeNet is selected as the teacher network.Through target distillation training,on the basis of giving full play to the small network detection speed,Finally,the ESPNet-KD track detection algorithm is proposed.Experimental results show that the detection rate can reach 63 FPS and the average crossover ratio of detection accuracy can reach 71.9%.It can meet the demand of detection rate and application of track detection in production environment.To sum up,the track line detection algorithm proposed in this paper effectively solves the current mine track detection accuracy and speed problems on the basis of considering the characteristics of the track environment,and provides a new idea for the final realization of the intelligent and unmanned mine transport system,especially the rail electric locomotive.Figure[25] Table[10] Reference[83]...