Research On High Speed Railway Foreign Invasion Detection Algorithm Based On Faster R-CNN

Non-track side objects such as pedestrians,vehicles,and animals entering the railway warning range will cause serious traffic accidents.Therefore,effective detection and identification of foreign objects entering the railway boundary is of great significance for ensuring the safety of railway traffic and pedestrian life.Traditional target detection algorithms use sliding windows to generate region suggestions,which has no pertinence,inaccurate prediction position,many redundant windows,large calculation and high time complexity.Using hand-designed features for feature extraction cannot adapt to the diversity of changes in target shape,detection scene and lighting,and has poor robustness.Traditional classifiers cannot process high-dimensional image data and need to be trained separately.The training steps are complicated and cannot meet the actual needs of the scene.The target detection algorithm based on deep learning has a faster detection speed than the traditional target detection algorithm,getting rid of the stereotyped artificial design features,allowing the machine to "learn" the potential feature connections contained in the image data,and achieve accurate detection of the target classification.In this thesis,a Faster R-CNN model with VGG-16 as the feature extraction network is proposed to detect foreign objects in railway intrusion,and it is adaptively improved to meet the needs of practical applications.This article mainly does the following research work:(1)Abandon the fully connected layer and use the global average pooling layer to complete the feature integration work.Since the emergence of convolutional neural network theory,adding several fully connected layers after the convolutional layer seems to become the standard configuration of the model.However,with the deepening of research,the shortcomings such as excessive parameter quantity,huge calculation amount,and the inability to effectively use the spatial position information between pixels gradually become prominent.The global average pooling layer can effectively reduce the model parameters while ensuring the performance of the model,so it is used as a substitute for the fully connected layer of the model to process the feature information of the last link.Using the models of global average pooling layer and fully connected layer respectively for training test and result analysis,it is found that the application of using global average pooling layer is better.(2)Increase the number of anchors in the regional proposal network.The original Faster R-CNN model uses 3 anchor scales and 3 aspect ratios for anchor parameter setting,which can generate 9 proposals.Considering the small feature of railway intrusion limit foreign objects in video surveillance,the small parameter setting of the anchor point is increased and the number of anchor scales is increased to four.Since different aspect ratios have less effect on the model,the three aspect ratio settings are retained.After the improvement,each sliding position can produce 12 kinds of proposals.By comparing the recall rate of the model to different specifications of targets under different anchor numbers,it shows that the improved algorithm improves the detection effect of small targets.(3)Use a model training method based on transfer learning.There is no special data set available in the field of railway foreign body invasion.However,the data set is also very important for model training,so the instance-based transfer learning method is introduced.Use the image data in the public data set similar to the feature space of this research field to supplement the existing data set in the field of foreign body invasion.Starting from "zero" to train a brand-new model is very expensive and the training results are affected by many factors.This thesis proposes to introduce a parameter-based transfer learning method,initialize the model parameters based on ImageNet,use artificially labeled railway foreign object infringement domain data sets to tune the model parameters,and complete the training of the improved model by alternating training.By comparing the accuracy of the models before and after using the transfer learning algorithm,the effectiveness of the training method proposed in this thesis is demonstrated.In order to verify the overall performance of the improved model,comparing the model with the original model and other deep learning models,it is found that the algorithm in this thesis can maintain a certain detection accuracy and has a faster detection speed.The accuracy of detecting foreign objects in railway intrusion reaches 97.81%,which has certain advantages in existing algorithms in the same field.