Multi-lane discontinuous lane line instance segmentation based on discrimination loss function

With the emergence of new technologies such as 5G networks and data centers, traditional infrastructure such as transportation, energy, water conservancy and municipal government have the basis for intelligent technological transformation. Smart high-precision maps are an important part of smart transportation and autonomous driving. That our automatically lane line segmentation provide technical support for high-precision navigation maps. There are two kinds of lane line segmentation methods currently, one is segmenting all lane lines in the scene semantically, and then uses post-processing for single-lane clustering, and the second uses the semantic or instance segmentation to split the pseudo-pixel-level continuous lane lines. For the instance segmentation of non-continuous lane line is still to be further studied, this paper proposed a LaneNet+ network to achieve lane line (including non-continuous lane line) instance segmentation end-to-end, this paper draws on the LaneNet neural network architecture, builds multitasking networks, semantically split network and embedded space network, respectively, and then cluster each lane line and achieve lane line instance segmentation. The total loss function is the arithmetic average of the semantic segmentation network loss function and the embedded space network loss function. The training sample dataset is a pixel-level labelled photos taken by the car’s dashcam, with a sample size of 12k, a number of iterations of 90,000, and the model was trained on the server of 2 Nvidia GPUs for approximately 12 hours. The model accuracy rate is 95%. Lane line instance segmentation can accurately distinguish the lane lines of different lane numbers, after that converted to the geographic coordinate system, and combine the GPS positioning information of the vehicle, then lane-level positioning and navigation of the vehicle can be realized.