South Africa as a developing country has to make the most out of the infrastructure that are available. Given the high level of crash involving pedestrians, it is critical that all means available are utilised to characterise pedestrian movements on the highway and pedestrian bridges. This paper will focus on using the existing camera infrastructure, but will extend its use to automatically detect and count pedestrians that use the pedestrian bridges. The pedestrian movement data can be used to aid with the evaluation of pedestrian safety campaigns, or to recognise trends in pedestrian movement. The paper presents the impact of various parameter changes to the state of the art technique used, as well as orientation suggestions for future installations. This is done to make optimal use of existing infrastructure, and provides an alternative to existing high-end systems. The methodology includes training a computer vision-based algorithm to recognise and count pedestrians for specific scenes, for example pedestrian bridges. The paper evaluates different suppression techniques to reduce false positives. The results show that 72% of pedestrians can be detected (a hit rate of 72%), with the camera facing a pedestrian bridge squarely from the side, so that silhouettes are clearly visible. High end products not using existing infrastructure typically have a hit rate of 70%-90%. The solution in this paper competes with high-end products, and can be expanded for infrastructure security applications, e.g. monitoring copper cables or monitoring of high risk areas.
Paper presented at the 35th Annual Southern African Transport Conference 4-7 July 2016 "Transport ? a catalyst for socio-economic
growth and development opportunities to improve quality of life", CSIR International Convention Centre, Pretoria, South Africa.