The effect of visual pavement marking properties on deep neural network detection performance

In this paper we discuss the effect of traditional day-time pavement marking properties on the performance of CNNs in the area of autonomous vehicles. In particular we examine the effect of pavement marking contrast on CNN performance. Similar studies have been undertaken in the past but most use proprietary hardware where the CNN output cannot be examined in detail. Here we train our own semantic segmentation CNN models on the LLAMAS pavement marker dataset and use the resultant probability maps as performance indicators. This also allows us to visualize the CNN filters and their activations. We found that contrast is a key measure in CNN performance. Other properties such as pavement marker area, orientation and luminance are also examined.