Abstract: The platform ends of the railway passenger station are non closed environment, which has the risk of illegal invasion.On the basis of explaining the principle of Faster-RCNN algorithm, this paper described the VGG16 model, RPN network and classified regression process in detail.The paper collected field data to make a sample set, and trained the station end intrusion detection model which could distinguish ordinary personnel, construction personnel and protection personnel.Five groups of experimental data with different parameters were tested and analyzed.It was determined that when the queue length of candidate area was equal to 300 and the number of recommended candidate area was equal to 15, it was the optimal parameter.The recognition accuracy rate of the model for ordinary personnel, construction personnel and protection personnel was 95%, 99% and 100%, the recognition recall rate was 97%, 99% and 100%, the average accuracy rate was 0.983 6, and the single frame detection time was 0.069 s.The results show that the algorithm can effectively detect ordinary personnel, construction personnel and protection personnel, meet the needs of real-time detection, and provide a new idea for platform end personnel intrusion detection.