Original title: An edge-adaptive vision-language framework for real-time open-set violence recognition in surveillance videos
Beibei Bian, Donghui Zhang, Jiamu Bai · September 2026
Security cameras need to spot violence quickly. Existing models struggle to do this on small, low-power devices. They use too much computer power, fail to track continuous video streams well, and cannot easily identify unfamiliar types of violent acts. To solve this, researchers designed a lightweight framework using a pre-trained vision-language model named CLIPA pre-trained model that learns connections between images and text from many internet examples.. It connects video imagery with descriptive text across two processing branches. This design helps the system adapt to new settings and recognize violence types it has not seen before. By updating only a small share of the model parameters, the system runs efficiently on low-power surveillance hardware. Tests across five datasets show it achieves competitive accuracy, making real-time public safety alerts possible on edge devices.
Imagine a security guard who checks people against a list of known troublemakers. But a new type of troublemaker appears that is not on any list. The guard would miss them. In this paper, instead of a fixed list, the guard has a general understanding of what trouble looks like, described in words, so they can spot anyone who fits a new description without retraining.
Most surveillance cameras have limited processing power and can only recognize a fixed set of violent behaviors. This system runs on those low-power devices and can identify entirely new types of violence by understanding language descriptions. That means public safety systems can be updated on the fly without replacing hardware.
The abstract notes 'preliminary explorations' for long footage, so temporal modeling may not be fully solved. Evaluations are on five datasets, but real-world surveillance conditions (lighting, occlusion) are not addressed. Competitive accuracy is reported without specific comparison to state-of-the-art, and deployment scenarios on varied edge hardware need further testing.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
An edge-adaptive vision-language framework for real-time open-set violence recognition in surveillance videos — Beibei Bian, Donghui Zhang, Jiamu Bai — Scientific Reports — doi:10.1038/s41598-026-72763-w
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.