CONG: Yet Another Human Tracking Dataset, but with a Little Secret
Yupaporn Wanna, Thanapong Intharah · 2024
We present a human tracking dataset, CONcealed Gun (CONG) dataset, where each human is identified as either a normal person, a person with a visible gun, or a person with a concealed gun. The dataset was mined from CCTV footages of robbery incidents uploaded to YouTube.com. A key distinction of the dataset is that we indicated the part of the video footage when the persons with concealed gun started to reveal the gun. So that we can use the part of the footage before the revelation as training data for people with concealed gun identification. Along with the dataset, we proposed a human with concealed gun detection algorithm which consists of two steps: human pose estimation step and human movement analysis. We explored the performance of the algorithm on different backbone models and footage lengths. We found that Random Forest on 2-second-tubelets achieved the highest recall at 59% with 32% precision, while Extreme Gradient Boosting on 1-second-tubelets achieved the highest precision at 32% with 57% recall. We showed that the CONG dataset a challenging dataset and we demonstrated that training a model to recognize a hidden trace is hard in terms of both dataset construction and the computer vision perspective.