Times Square: A Time Series Dataset for Semi-supervised Crowd Counting

Crowd analysis is pivotal for enhancing public safety, urban planning, and event management. Current crowd counting datasets, while useful, often fall short of capturing the entire scope of crowd behavior due to their time-independence. Moreover, using computer vision to forecast long-term crowd dynamics requires high costs associated with labeling extended time series of images. Thus, this study introduces a time series dataset gathered from Times Square in New York City, on which a semi-supervised learning technique is applied to enhance the efficiency of crowd counting. Spanning 16 days, our dataset comprises 4608 images captured every five minutes. A subset of only 576 images from the first two days was labeled. Our experiments evaluate the effectiveness of three semi-supervised crowd-counting models: DACount, MT, and MTCP, with different training ratios of labeled data (5%, 10%, 20%, and 40%). Results demonstrated the effectiveness of all three models with minimal labeled data. The DACount model consistently outperformed other models, achieving superior performance with only 5% of labeled data. Using the DACount model trained on 5% labeled data (28 labeled images), the entire time series of 4608 images was extracted. Results showed strong consistency with the manually labeled first two days, validating the ability of the semi-supervised models to perform well with a limited set of labeled data.