Soft-CSRNet: Real-time Dilated Convolutional Neural Networks for Crowd Counting with Drones

摘要

In recent years, the measurement of crowd density in a real-time video sequence has been a significant field of study. The use of these methods to stop protest scrambling, and social distancing to protect from COVID-19 is a crucial task nowadays. In this article, we introduce a different model for estimating crowd density based on front and vertical drone video sequences. Our proposition consists of an optimized version of a widely used crowd counting model called “CSRNET”. The proposed “SOFT CSRNET” is composed of two parts: a CNN front-end and CNN back-end. The front-end is composed of VGG16 layers constructed in the same way as CSRNet. On the other hand, in the back-end we select five convolutional layers of different size in the aim to get better results in less time. The results demonstrate that our method outperforms CSRNET in terms of MAE, image par second (ips) and proof of efficiency for a real-time videos sequence of drones. Our results are validated, executing the proposed method on Visdrone2019-DET and Visdrone2020-DET datasets.