Detecting and Preventing Drowning Accidents using Thermal Cameras

Olsson, Jonas Lundgaard ; Bonderup, Søren Arthur Bojesen

4. term

Vision, Graphics and Interactive Systems, Master

2016

2016-06-02

170

Siden 2008 er 8 personer druknet i Limfjorden ved Aalborg. For at mindske risikoen ønsker Aalborg Kommune at opdage hændelser med termiske kameraer. Denne afhandling undersøger, hvordan computersyn kan bruges til automatisk at registrere, når en person falder i vandet, og at advare på forhånd. Tre termiske kameraer er installeret ved havnen. Vi har analyseret, udviklet og testet et automatiseret overvågningssystem, der sporer personer i videostrømmen og registrerer fald i vandet. Systemet udtrækker automatisk en persons position i hver videoramme og følger bevægelsen gennem scenen med et Kalman-filter, en matematisk metode der glatter og forudsiger bevægelser. For at opdage fald bruges en virtuel trip-wire (en digital linje i billedet, der registrerer passage) og optisk flow, som måler bevægelse mellem billeder. I tests registrerede falddetektoren 100% af alle fald med i gennemsnit 0.08 falske alarmer i timen. Ud over falddetektion udviklede vi også en faldforudsigelse, der kan advare en operatør, før en person når vandkanten; i tests forudsagde den 23.67% af de kommende trip-wire-aktiveringer. Til udvikling og evaluering blev der optaget 155 timers termisk video i nattetimerne. Personers bevægelsesforløb blev annoteret i 56 timer til træning og delmodultests; de resterende 99 timer blev brugt som en accepttest. Resultaterne tyder på, at termiske kameraer kombineret med computersyn kan opdage og i et vist omfang forudse farlige situationer ved havnen om natten.

Since 2008, 8 people have drowned in the Limfjord at Aalborg. To reduce this risk, Aalborg Municipality aims to detect incidents using thermal cameras. This thesis explores how computer vision can automatically detect when someone falls into the water and provide an early warning. Three thermal cameras were installed at the harbor. We analyzed, developed, and tested an automatic surveillance system that tracks people in the video feed and detects falls into the water. The system automatically extracts a person’s position in each frame and tracks their movement through the scene with a Kalman filter, a mathematical method that smooths and predicts motion. To detect falls, it uses a virtual trip-wire (a digital line in the image that registers crossings) and optical flow, which measures movement between frames. In tests, the fall detector identified 100% of falls, with an average of 0.08 false alarms per hour. Beyond detection, we developed a fall predictor that can warn an operator before a person reaches the water’s edge; in tests, it predicted 23.67% of upcoming trip-wire activations. To build and evaluate the system, 155 hours of thermal video were recorded during night hours. People’s trajectories were annotated for 56 hours for training and module testing; the remaining 99 hours were used as an acceptance test. These results suggest that thermal cameras combined with computer vision can detect—and to some extent anticipate—dangerous situations at the harbor at night.

[This abstract was generated with the help of AI]

Drowning ; Vision ; Accident ; Aalborg Harbor ; Optical Flow ; Tripwire ; Kalman ; Tracking ; Pedestrian

Download
View record in AAU Student Projects