We address the problem of automating the previously manual inspection of asphalt distress and damage, which is time-consuming and subjective. With the help of a semantic segmentation neural network, we automate this process and make a pixel perfect detection. Our method is based on a previously published approach, validated for detecting distress symptoms in asphalt surfaces, for highway road data. The main challenge with the previous approach was, that the original training material did not yet contain significant samples of classes like foliage, drainage and additional types of road surface damage relevant to smaller secondary roads. Upon trying to retrain the underlying neural network it was discovered that methodological improvements were necessary to accurately detect these new classes. For now, we tried a new combination of a momentum-based Adam-Optimizer and a Cosine based learning Rate scheduler with restarting epoch also called Cosine Annealing. In theory, this should evade local minima of the error function, compared to the previously used polynomial decay without restarting epoch. The combination of both should yield a better minimum of the error function and higher accuracy rate.
Microsimulation is an essential tool in the planning and evaluation of large infrastructure projects or the implementation of new traffic rules for the inclusion of new road users. Efficient microsimulation on a citywide scale is provided by several software solutions, however, the representation of behaviors of unmotorized vulnerable road users (VRUs i.e., pedestrians, cyclists, scooter users) with their many degrees of freedom (including suddenly stepping on the road, crossing red lights) and their close interactions due to lack of space on sidewalks and other areas tend to be highly simplified. This work utilizes camera data collected in a busy road environment with several close interactions of all types of urban road users, to define decision-making rules and models of vulnerable road user behavior to be used in microsimulation. The simulation models are based on the observed interaction statistics, through the use of Surrogate Safety Measures (SSM, in particular Time-to-Collision-(TTC-), Post Encroachment Time- (PET-)), and the assessment of the kinematic behavior of interacting users in potential traffic conflicts.