Computer vision based pedestrian trajectory analysis
Samuel Richard Hislop-Lynch · The University of Queensland · 2018
Micro-level pedestrian simulation has the potential to provide a deep insight into the behaviour of people in public spaces. This allows researchers and practitioners to experiment with improvements to these spaces which can lead to more efficient, less stressful journeys for pedestrians. However, in order to simulate the behaviour of a public space, a great deal must be known about the existing conditions. Herein lies the problem. In all but the most trivial cases, gathering a sufficient quantity of information in order to calibrate a simulation can be extremely time-consuming and error prone. Currently, researchers and practitioners have four broad approaches for collecting this information: human-based observation, point-counting devices, human-based video analysis and computer-vision based video analysis. Here, this research discounts the possibility of utilising electronic positioning data such as those generated by Bluetooth and Wi-Fi. This is due to the difficulty that researchers and practitioners may have in acquiring such a dataset and that passenger profiling cannot be easily undertaken with this technology. Human-based data collection approaches suffer from speed and accuracy issues, although, they also provide the greatest level of detail. Point-counting devices, on the other hand, offer higher levels of accuracy, but lack the ability to provide finer levels of detail. An alternative option, computer-vision based video analysis, can offer practitioners and researchers the best of human-based methods and point-counting devices without the drawbacks. However, implementing a pedestrian trajectory analysis tool is a daunting task and few are willing to outlay the time required to fully understand the prerequisite computer-vision field. To this end, this research has delivered a framework and an associated library which greatly reduces the barrier to entry into the field of computer-vision based pedestrian trajectory analysis. The framework, presented here, can be used as a template for other practitioners and researchers who wish to build their own tools in any language. This is targeted at those who can’t afford to invest many months in the pursuit of computer vision knowledge. Furthermore, a C++ library, developed alongside the framework, provides programming primitives that can be used to quickly build a high frame-rate pedestrian tracking and trajectory analysis tool, without having to start from scratch. As a result of this research, the barrier to entry into the field of computer-vision based pedestrian trajectory analysis has been greatly lowered.