Pedestrian Intruding Railway Clearance Classification Algorithm Based On Feature Fusion

With the rapid development of Chinese high-speed railway,the importance of railway transportation in personal safety is increasing,and the safety issue after the operation of high-speed rail has become a top priority.The objects intruding high-speed railway clearance will lead to traffic accidents,seriously endangering state property and personal safety.For the problem of the existing railway image object intruding detection system can only detect the alarm image,it is impossible to distinguish whether the correct alarm is caused by human intrusion or the false alarm caused by light interference.This thesis aims to develop a set of improved convolutional neural network pedestrian classification algorithm based on fusion features to correctly separate the pedestrians in the alarm image,which reducing the false alarm caused by light interference and solving the problem of high false alarm rate of the objects intruding railway clearance detection system.In order to accurately distinguish the alarn pictures from pedestrians or light disturbances,this thesis proposes to improve the high-level Alex features extracted by the convolutional neural network and the HOG features,LBP featUres for multi-featured weighted fusion,and classify the fusion features using fully connected classification network.First,collecting the alarm pictures of the railway scenes of the existing railway image object intruding detection system,and building the pedestrian intruding railway clearance sample database one by one by using the pedestrian-free standard,the constructed sample library includes 5240 false alarm samples and 2216 correct samples of pedestrians.Subsequently,the HOQ LBP single feature and weighted fusion feature of the image are extracted separately,input into the SVM for classification and recognition,and the parameters of the SVM are optimized.The experimental results show that the recognition accuracy of fusion features is significantly higher than that of single features.However,due to the complexity of the railway scene,the interference of environmental factors and the insufficient number of samples,the accuracy of SVM algorithn based on traditional features does not meet the system requirements.Finally,this thesis introduces a convolutional neural network and proposes to weight traditional featiires and the high-level features extracted by the convolutional neural network as the fusion features to improve the accuracy of the classification algorithm,then uses the deep separable network and the LI norm cropped convolution kernel to streamline and optimize the network to improve real-time performance.The algorithm reduces the running time and calculation amount of the algorithm,and can accurately and quickly distinguishes the alarm picture from pedestrian or false alarm caused by light interference.This thesis proposes a fusion feature classification algorithm based on improved convolutional neural network to make up for the lack of existing methods.The experimental results of the pedestrian intruding railway clearance library show that the algorithm proposed in this thesis has an accuracy of 98.58%for 1472 test sample images and the test time of 10.86 ms for single image,which greatly improves the accuracy of pedestrian classification detection and reduces false alarms caused by nighttime light interference.This algorithm has strong applicability and real-time performance,and is of great significance for reducing system false alarm rate and ensuring railway safety.