View Presentation
Cycling encompasses many societal benefits. It influences a community’s safety, economy, environment, equity and health. The number of cyclists on the roads is highly influenced by their perception of safety. To determine road safety, it is fundamental to have a common metric, so that risk factors can be determined and compared. Using Google Street View (GSV) imagery is a cost-effective approach to analyse urban environments. Due to the high number of images needed to extract accurate results, models to automatically detect objects and structures are used.
CycleAI aims to detect the areas of higher perceived risk for cycling worldwide. This project used object detection and image segmentation models to extract cyclists’ road risk factors from GSV images of Lisbon. This involved analysing a GSV dataset, before using two state-of-the-art tools, YOLOv5 and NVIDIA, to detect objects and segment images, respectively, and further analysing their results; determining the limitations of YOLOv5, NVIDIA and suggesting ways of making cyclists’ safety assessment more accurate.
Approximately, 4500 objects were identified, and 400 million pixels labelled on the 1000 images from the Perception Poll. Cars (84%), people (7%) and trucks (3%) were the most common objects detected. Sky (22%), road (21.8%), buildings (19,6%) and vegetation (13,4%) the most present structures.
Perception Poll allowed to weight the contribution of all objects and structures to the perception of safety of thousands of bicycle riders. More than 17 000 votes from 26 different countries were collected.
From the analysis of these images and the respective safety scores obtained from Perception Poll, CycleAI was able to statistically correlate cyclist’s safety perception with specific objects and structures. Significant positive correlations between trains, buses, cars (strongest) and low safety scores were found. On the other hand, bicycles and people (strongest) tend to be associated with the highest safety scores. Additionally, a neural network was trained with the capacity to predict a safety score for a road image with 70% accuracy.
Future directions include increasing the availability and resolution of GSV images. Train YOLOv5 and NVIDIA with datasets containing a higher number of categories relevant for road safety. Define a safety metric to weight and combine (at a road level) detected objects or segmented structures. Finally, to process street view images or video in real-time would allow to better capture the dynamics of road safety (video). In CycleAI, we are compromised in holistically tackling these issues.