Three‑Path 3D Network: Fast Video Crime Spotting Made Simple

A new approach called TriPath3DNet turns short security clips into quick, accurate alerts. It uses a lightweight 3‑layer “ResNet” model that watches the video in three ways: short bursts of motion, long‑term background clues, and quick frame differences. These three lenses help the system notice subtle problems like someone loitering or sneaking into a forbidden zone, even when the light changes or people block each other. The team tested it on four sets of real footage, including two new ones named Virat1‑RC and Virat2‑RC. TriPath3DNet scored better than many popular methods, reaching up to 95 % correct guesses and a 99 % area‑under‑curve score. It can process each 50‑frame clip in about 130 ms, which means it works almost instantly on a standard GPU and uses only about 33 million parameters.In tough data sets where other models stumble, TriPath3DNet shines. For example, on the UCF‑Crime set it outperformed transformer‑based rivals that usually lag behind. Even when a competitor had a slightly higher overall accuracy on one data set, TriPath3DNet delivered a much sharper detection rate for real threats. A study of each part of the model showed that all three paths add value. Visual tools like Grad‑CAM confirmed that the network focuses on the right spots in both space and time, giving developers confidence it is not just guessing. By designing with real‑world cameras in mind—fast, low‑memory, and able to run on edge devices—this work moves the field closer to everyday use in factories, malls, and public spaces.