Cycle way condition evaluation using Lidar and deep learning

The Austrian part of the Danube River, on its full length, is accompanied by towpaths, historically used by men or working animals to tow boats. Today, these paved towpaths are used as cycle ways and are part of the EuroVelo EV6 “The Rivers Route” at a length of ~400 km. Near Vienna, more than 500.000 cyclists are counted each year on the cycle way. Furthermore, the towpaths ensure the accessibility of the riverbank for emergency services, construction works, etc. With that in mind, viadonau, maintainer of the towpaths, decided to establish a management system for the maintenance of that network. This includes a yearly condition survey. Together with via donau, AIT established a survey and evaluation scheme, covering two aspects: structural condition and ride comfort/ride safety.
The survey collects panoramic images of the cycle way corridor and Lidar data from the surface. Using a profile scanner, a width on the ground of 4.5 m is covered. From the raw Lidar data, high resolution 3D point clouds and 2D intensity images with a section length of 5 m are generated. As the width of the paved cycleway is usually below 4.5 m, in a first step a pre-trained convolutional neural network (CNN) is used to discriminate pavement from vegetation. Only the paved parts of the 2D images and the 3D point clouds are considered for further evaluation.
For the structural evaluation, cracking, scaling/spalling and potholes are detected by the CNN on the images. Crack lengths are multiplied by a width 0.5 m to calculate the affected area. Scaling, spalling and potholes are detected as surface areas. Using different weights for crack area and scaling/spalling/potholes, a percentage of the area affected is calculated. This percentage is transformed to a five-stage dimensionless scale.
For the ride comfort/ride safety evaluation, the point cloud of the remaining pavement is divided into tiles of 20x20 cm. For each tile the spread in z-direction is calculated. The spreads of all tiles are aggregated using the 95 % percentile for each 5 m section. The spread percentile is then transformed into a dimensionless evenness indicator. The evenness evaluation focuses on short-wavelength irregularities like root heaves and steps that present a fall risk for cyclists and reduce ride comfort.
This evaluation approach ensures that the full width of the pavement is evaluated for structural and ride comfort/ride safety deficits. It equally covers the maintainer’s needs and user’s comfort and safety needs respectively. Taking both indicators into account, the appropriate maintenance measures can be derived for each section.