Research And System Design Of Rail Obstacle Detection Method Based On Deep Learning

With the continuous increase of railway operation mileage in my country,the railway coverage has become very broad,which makes the driving environment more complex.At present,the detection of track foreign objects is mainly based on manual inspection and traditional detection technology.These detection methods have shortcomings such as high cost and low adaptability,and cannot meet the needs well.In this paper,a deep learning-based foreign object detection method on rails is proposed,and the one-stage algorithm YOLO-v5 is selected for the detection of foreign objects on rails,aiming to improve the detection accuracy and detection speed.The main research contents of this paper include.Firstly,through conventional processing such as image denoising,averaging,and flipping,as well as image enhancement and expansion operations,the production of orbital foreign object image sets is completed.Secondly,YOLO-v5 network s-version Although the detection speed fast,but the performance of detecting small targets is poor.In this paper,an attention mechanism is added to the CSP module of the network to enhance the feature extraction ability and rich feature information,and CIOU is used to replace the GIOU loss function.Thirdly,the difference between some orbital foreign objects and background features is not obvious.if the size of the foreign object changes,not only the detection accuracy will decrease,but also the problem of missed detection or false detection may occur.In this paper,the Bi-FPN feature fusion method is used to improve the network and improve the ability of feature information fusion.Fourth,this paper uses QT library designed a set of rail foreign object detection system,and completed the test of the rail foreign object detection system by loading the weight model and selecting test pictures and videos.The experimental results show that,through the analysis of the results of adding CA and CBAM mechanisms to the basic network,the training loss curve with the CBAM mechanism has a faster convergence and decline trend,and its model accuracy and mAP value are better than those with the CA mechanism.Compared with the detection network before the improvement,the decrease of the target loss curve becomes steeper,the curve convergence speed is better,and the mAP value is increased by 1.1%.After verification by the test set,the recognition accuracy rate is increased by 0.92%.If the size of the target changes,the missed target can also be identified.The server used this time to detect each image takes about 0.291ms less than before the improvement.Therefore,this paper can further improve the detection performance of the model through some improvement methods.