Railway intruding pedestrian image sequence generation method with pose transfer techniques

Pedestrians intruding detection on deep learning method is vital for the safety of high-speed railway operations. The accuracy of deep learning object recognition is significantly depended on the volume and diversity of training samples. Addressing the issue of insufficient and hard-to-obtain intruding images in high-speed railway scenes, this paper introduces a novel method for generating railway intruding pedestrian image sequences with pose transfer. This method can generate necessary intruding samples and produce the annotated files automatically for deep learning training. To tackle the challenges of pose feature extraction, a pose transfer module based on generative adversarial network is proposed; furthermore, a feature fusion module based on spatial attention mechanisms has been proposed to preserve essential pose features. Experimental results on the DeepFashion dataset demonstrate that the proposed modules significantly enhance the quality of generated pedestrian images and the accuracy of pose transfer. In the same data set, compared with the classical pose transfer algorithm, the algorithm in this paper improves the SSIM index by 5.6% and the pckh index by 2.9%. Finally, the effectiveness of the intruding image generatio n is also verified by the improvement of detection algorithm accuracy.