CBL - Campus del Baix Llobregat

Projecte llegit

Títol: Detección automática de violencia en vídeo para vigilancia con drones

Estudiants que han llegit aquest projecte:

Director/a: VALERO GARCÍA, MIGUEL

Departament: DAC

Títol: Detección automática de violencia en vídeo para vigilancia con drones

Data inici oferta: 03-02-2026     Data finalització oferta: 03-10-2026


Estudis d'assignació del projecte:
    GR ENG SIST AEROESP
Tipus: Individual
 
Lloc de realització: EETAC
 
Paraules clau:
Detección de violencia, Redes neuronales, Visión por computador, Drones, Vigilancia
 
Descripció del contingut i pla d'activitats:
 
Overview (resum en anglès):
This Bachelor's thesis studies the feasibility of using automated surveillance drones, equipped with a video analysis model capable of detecting violent situations, as a support tool for security forces. The aim is for the operator to receive an alert only when the system detects a possible assault, instead of having to watch the footage continuously. For example, the drone could be deployed from a patrol car to cover nearby streets that are not directly visible from the vehicle.

For the detection task, a model based on a pretrained ResNet18 neural network has been developed and trained. It extracts the features of each frame and combines them through temporal averaging, followed by a custom classifier. The model was trained on the public AIRTLAB dataset, complemented with self-recorded videos, through an iterative process with validation on recorded videos and real-time validation. The system integrates a full pipeline of capture, analysis, alert with the corresponding clip and storage in a database, together with a ground station (dashboard) that communicates through MQTT and WebRTC.

After comparing six configurations, the one that best balances the reduction of false positives on non-violent actions with an adequate detection of violence has been selected. The result is a functional prototype, although with a generalization ability limited by the size and low variety of the dataset.

Finally, a feasibility study of a real deployment on a drone has been carried out, analysing the communication link, weight, autonomy, cost and the legal and environmental aspects. Autonomy is the main limiting factor, and sustainability is conditioned by battery wear. It is concluded that, although the concept works at a prototype level, its real deployment is not currently viable, mainly due to the limited flight autonomy, and would require expanding the dataset, especially with real aerial imagery, and significantly improving the drone autonomy.

© CBLTIC Campus del Baix Llobregat - UPC