Analysing pedestrian movement with the Flowity software

Abstract: This thesis surveyed existing technologies and research on pedestrian detection systems using video as input. The purpose of the study was to investigate what such a software would need to produce to be relevant for the fire safety engineer in outdoor circulation- or egress applications. The thesis tested a newly developed detection software called Flowity. The software, developed by ÅF Digital Solutions AB (a subsidiary to AFRY AB) utilizes machine learning and artificial intelligence algorithms to identify and detect objects and pedestrians. The objective was to establish a list of factors that the Flowity software should be able to extract and how, if to be useful for the fire safety engineer. The objective was also to conduct a case-study test of the software on a video of pedestrian movement and to evaluate/compare the capabilities to accurately identify and quantify pedestrian movement with the Flowity software and do a comparison to manually collected data. Results from the study show that the Flowity software could identify people and automatically presented them as detected pedestrians with a detection accuracy of 70 %. This applies for this case study, which was an outdoor highresolution video recording containing 576 pedestrians, with the camera placed 4 meters above the walking area. The software managed to provide data on movement patterns, route choices of detected pedestrians as well as measuring movement speeds, flows and densities at different sections. The maximum global people density measured with the software was 0,13 persons/m2 and the maximum local density was 3 persons/m2. When comparing the manually and software measured flows and densities, there was no statistically significant difference between the measurement methods. However, a comparison between manually and software measured movement speeds showed a statistically significant difference between the measurement methods. A 14,2 % higher average flow was measured with the manual counting and a 15,1 % higher average global denisity. The software measured a 32 % higher average speed than what was manually measured. Uncertainties connected to the manual measurements and unknown influence of factors on detection performance might have impacted the results.