SSD Algorithm And Its Application In The Abnormal Object Detection Of Railway Scenes

With the increasing density of railway operating networks,the safe operation of railway systems has become particularly important.Effective measures are needed to minimize the impact on railway safety operations.It has become an important issue in railway operations that researching reliable and accurate algorithms to ensure railway transportation safety today.When the traditional target detection algorithm is applied to the detection of abnormal targets in railway scenes,it has the disadvantages of high feature extraction difficulty and the performance of the algorithm being vulnerable to the external environment.In recent years,the feature extraction of objects using convolutional neural networks has 'gradually replaced the artificial target feature design and has good robustness.This topic researches SSD(Single Shot MultiBox Detector)algorithm with high real-time good performance and detection accuracy performance,and conducts its application in foreign object detection for railway scene.The research collects the object images of the railway scene and the common scenes containing the categories to be detected.Combined with the idea of transfer learning,the SSD network is trained in different ways,and the model for the detection of foreign objects in the railway scene is obtained and the performance of the model is analyzed.Aiming at the problem that the SSD algorithm has poor detection performance on small-sized targets,a multi-block SSD algorithm is proposed.The process firstly divides the four overlapping areas of the detected image,and then performs the SSD network detection on the divided pictures.Finally,the detection result is integrated and filtered,and the detection performance of the traditional SSD detection algorithm in the small target object category is improved.The experimental results show that the SSD algorithm can complete the task of detecting foreign abnormal objects in the railway scene.In the Python 3.6 version of the Anaconda3 environment,the performance evaluation data of muilti-block SSD networks is obtained.According to the object detection confusion matrix,it can be calculaed that the mFl of the muilti-block SSD network up to 96.6%,which is 9.2%higher than the traditional SSD model.Besides,the F1 of the stone category is 23.2%higher than that of the traditional SSD network.From the comparison of the final picture results,it can be found that the improved multi-block SSD algorithm can detect more targets and the position of the object box is more accurate.