一种基于无人机拍摄视频的人群计数方法

rowd counting based on uav videos is a challenging visual task, such as VisDrone2020 evaluation. To improve the performance of crowd counting network model, we propose a strategy to filter out building interference through semantic segmentation. Since VisDrone-CC2020 training data set does not have segmentation annotations, we present an effective semi-automatic segmentation annotation method, which greatly improves the annotation efficiency of manual labeling. In training phase, the HRNetv2 semantic segmentation network model is fine-tuned based on our labeling data. In testing phase, HRNetV2 network output is processed by morphological algorithms, and we get a building segmentation map with complete and smooth boundary for each scene in VisDrone-CC2020 test dataset. Then the building segmentation map is applied to our trained model to filter out building interference. At the same time, an optical flow extraction module is added in the crowd counting network to provide target motion information. This crowd density prediction scheme was applied to visdrone-CC2020 international evaluation, and ranked ninth in the evaluation.