Lane Marking Segmentation in the Autonomous Driving Scenario with Deep Convolution Neural Networks

This thesis represents flatly different approach on Lane Marking Semantic Segmentation, aiming at improving seen results, and ending up with compeletly new techniques. DeepLabV3 Plus [7], extending DeepLabV3[5], is trained on ApolloScape dataset. Large-scale dataset contains a diverse set of stereo video cropped sequences, recorded in street scenes from different cities, with high quality pixel-level annotations of 110 000+ frames. Xception network [8], being extension of InceptionV3 network[8], is considered on such a particular research, together with DeepLabV3 plus. Proposed solution has variety of advantages in the task of Semantic Segmentation, in terms of providing, new techniques such as encoder-decoder network, Spatial Pyramid Pooling with Parallel Atrous Convolution layers[7], Depthwise Separable Convolution as well as Multi-Grid Method which all are broadly discussed in this thesis. Regardless of several state-of-art methods, there are possible challenges, needed to be explicitly analyzed, and taken good measures, to propose superior achievements. Confronting difficulties in Semantic Segmentation, concerns to obtained results, related to mIoU and Class Accuracy of 38 classes, which are, in turn, caused by Cross Entropy Loss for unbalanced dataset and Random Scale Crop function which operates on randomly scaling ground-truth images, resulting disappearance reasonable information on images, on the other hand, having 38 classes on images, bring challenge to network to classify and semantically label, great variety of road signs on the images.there are experimented methods suggested for those ongoing issues, for example, replacement of Cross Entropy loss, with Weighted-Cross-Entropy Loss, Random Scale with Standard Random Crop, and deployment of Center Crop technique and training with two classes, namely, ”Lane Marking” and ”Non Lane Marking”, all in all have dramatically improved previous outcomes, in particular, with advent of Weighted Cross Entropy Loss.