Currently, online gaming is continuously gaining popularity, largely due to the easy accessibility of computer devices and convenient internet connections. There are various types of online games, such as multiplayer first-person shooters like Counter-Strike: Global Offensive and Valorant. Competing against players worldwide in professional-level tournaments is known as the "E-sports" or Electronic Sports scene. In E-sports, players demonstrate their skills and expertise in competitive gaming at a high level. The growth of the E-sports industry has been remarkable, impacting both the sports industry and global society. This growth has led to professional-level competitions receiving support and investment from major companies, as well as garnering interest from players and viewers worldwide. One crucial factor that gives teams an edge in competitions is data regarding players' gaming performance. This data, analyzed by players and experts, serves as a significant variable in team training, increasing their chances of winning against opposing teams.
In today's gaming landscape, many games like CSGO or DOTA 2 provide replay files and log files that publicly disclose game data from start to finish for download. However, in games like Valorant, which researchers may analyze, there are no publicly available replay files or log files for download. As a result, analysis typically involves scrutinizing gameplay recordings from videos, which can be time-consuming or less efficient in extracting comprehensive insights.
Researchers are interested in developing an analysis system for the game Valorant from gameplay recordings, which can categorize actions into buying equipment, shooting, planting, using skills, defusing bombs, collecting orbs, dying, and class switching. Since interesting events often revolve around actions, this system aims to efficiently help users analyze player actions and identify events that impact player actions. The researchers will develop the system in the form of a web application. For this development, a dataset will be created from 100 Twitch videos featuring characters such as Harbor, Jett, Yoru, Sage, and Sova, with a total video length of 56 hours, 51 minutes, and 70 seconds, and a total video size of 410.70 GB. The researchers will choose the MoViNet A5 model for development, which will be divided into five models with the following accuracies: Harbor A5 model with 97% accuracy, Jett A5 model with % accuracy, Yoru A5 model with % accuracy, Sage A5 model with % accuracy, and Sova A5 model with 92% accuracy. Regarding the web application, the system will start by inputting the video into the web application process, which the researchers have implemented in the backend to separate the eight actions from the video. The resulting output will be all the actions that occur within the video, displayed on the web application interface and downloadable as a CSV file.
The main benefits of this research include providing guidance for others to develop models for gesture recognition and allowing Valorant players to utilize the Valolyze system to analyze Valorant gameplay from gameplay recordings.
