Fight Detection for a Video Analytics System

Task

The client approached us with the goal of creating a video analytics system capable of identifying dangerous behavior, such as fights, in public places. Since there were no open datasets with such content, the client needed to rapidly collect videos of fight scenes in public and indoor locations. Given our ongoing partnership with the client, they trusted Unidata to handle this complex project.

Solution:

1. Comprehensive Data Collection:
   To fulfill the project requirements, we needed to collect 200 high-quality videos of fight scenes, shot in 4K with wide-angle cameras. We selected a variety of backgrounds and involved multiple actors for each scene.
2. Coordination with Local Talent:
   We developed the video scripts and cast actors from boxing clubs in Brest. The filming took place both outdoors and indoors, ensuring a diverse range of environments that the client could use for training their system.
3. Overcoming Weather Challenges:
   During the outdoor shoot, an unexpected snowstorm created significant challenges. Cold temperatures drained camera batteries quickly, so we used power banks to maintain consistent operation. We also had to regularly clean the cameras from snow and provide actors with hot beverages to keep them warm. Due to the weather, 10 videos required reshooting on the following day, but the project still stayed on schedule.
4. Optimized Workflow and Quality Control:
   Throughout the filming process, we conducted regular checks on video quality to ensure every detail was captured correctly. We utilized multiple cameras to provide different angles, which enhanced the final dataset’s quality.