Temporal Extension for Encoder-Decoder-based Crowd Counting Approaches

Published in 17th International Conference on Machine Vision and Applications (MVA), 2021

Crowd counting is an important aspect to safety monitoring at mass events and can be used to initiate safety measures in time. State-of-the-art encoder-decoder architectures are able to estimate the number of people in a scene precisely. However, since most of the proposed methods are based to solely operate on single-image features, we observe that estimated counts for aerial video sequences are inherently noisy, which in turn reduces the significance of the overall estimates. In this paper, we propose a simple temporal extension to said encoder-decoder architectures that incorporates local context from multiple frames into the estimation process. By applying the temporal extension a state-of-the-art architectures and exploring multiple configuration settings, we find that the resulting estimates are more precise and smoother over time.