Research On Semi Supervised Road Drivable Area Detection Algorithm

In this graduation project,We propose a semi-supervised road drivable area detection algorithm.The main application scenarios of the algorithm is autonomous vehicle.The semi-supervised algorithm we chose is based on pseudo labels and the base model is semantic segmentation neural network Deeplab V3+.The datasets we choose are semantic segmentation dataset Pascal VOC 2012 benchmark,drivable area detection dataset BDD100 k and the dataset of HIT road scene information.We propose three main algorithms to optimize the semi supervised algorithm.The dynamic region correction algorithm can make full use of the correction ability of neural network to correct the wrong region of the pseudo labels during the training.The weight optimization algorithm use a weight matrix to give a weight to the loss value of each pixel according to the output probability of the model.The model can adjust the learning ability of different regions according to the degree of confidence in different regions of the picture and the models give priority to learning their most accurate knowledge.Teacher network algorithm use a complicated teacher model with strong learning ability but low efficiency to make the pseudo labels.We use a relatively simple and efficient student model to learn the knowledge provided by the teacher model.It can improve the quality of pseudo labels while ensuring the running speed of the model.Experimental results show that our three semi supervised learning optimization algorithms can improve the effect of semi supervised learning significantly.With the semi-supervised road drivable area detection algorithm,the autonomous vehicle can continue to collect road information and it can use semi supervised learning algorithm to continuously learn the knowledge of the road that vehicles often travel to realize the lifelong learning of autonomous vehicle.