Research On Perimeter Intrusion Detection Algorithm Of High Speed Railway Based On Vision

According to statistics,by the end of 2018,the operation mileage of high-speed railway in China has exceeded 29000 km.With the expansion of road network and the increase of train speed,the perimeter intrusion detection of high-speed railway has been raised to a new height.At present,the perimeter intrusion detection of high-speed railway is lack of mature and reliable methods.It only relies on human monitoring,which not only consumes a lot of labor,but also leads to the problem of missing detection.Using the existing high-speed railway integrated video monitoring system to improve the level of perimeter security is an effective solution to the high-speed railway perimeter security.Because the camera is located in a high position and the distance between them is far(every 200m),there are a large number of small targets in the video,so it is difficult to detect these small targets with general image detection methods.In addition,real-time alarm is needed for perimeter intrusion.And the current target detection method can’t realize real-time alarm.It is difficult to improve the detection speed while ensuring the detection accuracy.In view of the above problems,the main work of this thesis is as follows:(1)Because the detection is carried out on video,the motion information between frames will be lost by using the target detection method on the image.So a fusion method of motion information and target detection is proposed.The moving branch uses the method of moving foreground detection.The target branch uses the two-stage target detection method based on deep learning.By fusing the detection results of two branches,we can judge whether the detected target is an intrusion target.It can not only effectively detect the intrusion target,but also eliminate the false detection of irrelevant moving objects or static objects.(2)There are a large number of small targets in the detection video,the general detection method will have many problems of missing detection.So an improved Cascade Mask R-CNN is proposed in the target detection branch.The model uses cascaded structure to get the accurate target location.At the same time,based on the original model,the multi-scale feature extraction module based on the feature pyramid network(FPN)and the spatial context enhancement module based on the empty pyramid aggregation(ASPP)sub network are added.The validity of the model is verified in the actual high-speed railway perimeter video data set.The results show that the model can achieve the railway perimeter intrusion detection in different scenarios.At the same time,the new model improves the F-measure of small target detection by 0.24 compared with the original model.(3)Aiming at the problem that the speed of branch detection is too slow to meet the real-time detection of railway perimeter,an acceleration algorithm based on optical flow is proposed.It uses a large number of redundant information between video frames.The improved algorithm determines a key frame every ten frames.It uses convolution operation to extract features on the key frame.It uses optical flow based feature warping algorithm on the non key frame,and warps the key frame feature map to the current frame to obtain the current frame feature map.Because the speed of optical flow and distortion is much faster than convolution operation,it can be accelerated.This method balances the detection accuracy with the detection speed,and greatly improves the detection speed at the cost of very small detection accuracy.