Research On Multi-scale Target Detection Algorithm For High-speed Railway Perimeter Intrusion

In recent years,China’s railway construction has achieved rapid development,and the mileage of railway operation has risen rapidly.Due to the long railway line,the great climate difference along the line,the limited perimeter protection facilities and other reasons,the incident of foreign body invasion occurred sometimes.The incident of foreign body invasion in railway not only hinders the safe operation of trains,but also threatens the safety of people’s lives and property.Timely and accurate discovery of foreign body invasion in railway is of great significance to ensure the safe operation of railway.However,due to the large difference of target scales in railways,the detection ability of existing detection methods for railway foreign bodies needs to be further improved.Therefore,a multi-scale feature fusion algorithm is proposed in this paper to improve the detection level of restricted foreign bodies in railways.In this paper,images of foreign body invasion along the railway were collected and made into a railway foreign body invasion data set.According to the definition of large,medium and small targets in COCO data set,this paper establishes the target classification standard of railway data set.According to the analysis of railway data set and common data set,the target scale distribution of the two kinds of data sets is very different.The target distribution of railway data set is more concentrated,more than90% of the targets are medium and small scale,while the target distribution of common data set is more uniform.Most of the existing backbone networks of railway foreign body intrusion monitoring algorithms are designed based on the target scale of common data sets,and there are some limitations in their application to feature extraction of railway foreign body intrusion targets.To solve this problem,this paper designs a new multi-scale feature fusion backbone network(MIDO network)for railway scenarios.MIDO network design includes multi-scale feature fusion module(MIDO module)design and backbone network connection design.MIDO module has the structure of multi-input and double-output,and the internal set of self-learning weighted fusion coefficient,so that the MIDO module can process more abundant feature information.Different MIDO modules are connected to form a MIDO network.The network consists of 4 layers with 6 modules in each layer.Each layer of the network outputs a group of feature maps.In this paper,channel weighting and feature weighting are designed to fuse the 4 groups of output features between the same scale and different scale.In the experiment,the detection accuracy of the algorithm reached 83.3% in the public data set VOC and 91.1% in the railway data set,both better than the current backbone network.In this paper,a railway foreign body invasion monitoring system is built based on MIDO network.The hardware of the system includes railway field equipment,connecting equipment and video processing equipment,and the software algorithm mainly includes accuracy improvement algorithm and speed improvement algorithm.The speed enhancement algorithm uses the region segmentation method.The image region is divided into sky region,area outside the ground limit and area inside the ground limit by region segmentation method.Different detection methods are adopted for different areas to speed up the algorithm.The accuracy improvement algorithm uses model switching method to combine the long-range specific model and the nighttime specific model with the general model to improve the detection accuracy of the algorithm for long-range targets and nighttime targets.In the system experiment,the detection accuracy of the system is 85.4%,62.7% and 31.8% for targets 0-50 meters,50-100 meters and more than 100 meters,respectively.The speed of the model is increased by 6% after the speed improvement algorithm.After the accuracy improvement algorithm,the detection accuracy of the model for distant targets is improved by two times,and the detection accuracy for daytime targets and nighttime targets is improved by 5.9% and 0.4% respectively.In the experiment,the system shows good detection accuracy and speed,and meets the requirements of railway intrusion foreign body monitoring.