Pedestrian Occlusion Detection Method based on Human Image Segmentation

Improving pedestrian detection accuracy in occlusion is a priority for developing intelligent transportation and autonomous driving. Traditional pedestrian occlusion detection methods still need to achieve a more satisfactory level of accuracy for pedestrian detection. This paper aims to study a new method to improve pedestrian detection accuracy in an occlusion environment. The core idea of this method is to regard the non-portrait part of the detection object as interference, use portrait segmentation as a preprocessing method to remove the non-portrait part, and only keep the portrait part. After recalibrating the training and test data of the model, YOLOv5 is used for model training to obtain a higher-level recognition model. This paper uses real-life pedestrian data CVC-05 obtained from natural scenes for testing. The map0.5:0.95 obtained in this paper reaches 0.7739, which is significantly higher than other algorithm models and is of great significance to the field of automatic driving pedestrian detection.