چكيده به لاتين
By removing the driver, autonomous vehicles need many sensors in order to achieve a correct perception from the environment. Machine vision based on deep learning is one of the most common ways to detect vehicle’s trajectory and obstacles. However, when faced with lots of domain shift between train and test datasets, this sensor’s performance decrease drammatically. Having this challenge in mind and by reling on the machine vision and deep learning, vehicle’s trajectory and obstacles in this work is detected. At first, aim is on the generalization of the deep learning models in the field of object detection by training the Faster-RCNN and SSD. There are 12.72% and 14.52% improvements in these models performance by defining new train scenarios in the mean average precision (mAP) metric compared to other works. Also in the next step, model’s robustness is subject to a huge challenge by applying different noises on the test database. Results show that our strategies imporved Faster-RCNN Resnet101’s accuracy by 6.85%, 7.07%, 6.96% and 10% in s1, s4, mPC and rPC metric respectivelly. Simulation and experiments in cross-domain evaluation show that our strategies caused 14.56% and 5.5% improvements in mentioned model’s performance in real and simulated images respectivelly. Applying semantic segmentation for detecting vehicle’s trajectory and obstacles simultaneously in pixel level is another goal of this thesis and in this field we take advantage of fast and state-of-the-art networks like DABNet, ContextNet and FastSCNN. Also in order to answer time consuming challenge to prepare manual annotation, this thesis uses auto-generated annotation with teacher-student approach. With respect to the extracted results in model’s evaluation, there is only 1.2% performance difference compared to the same model which is trained after spending very long time only to prepare manual annotation by humans. Also, it is proved that introducing unstructured environments to model end up with 5.1% and 1.9% improvements in teacher and students accuracy respectivelly. Finally, there is 1.3% improvement in FastSCNN model which uses the combination of manual and auto-generated annotations compared to the same model which only trained by humans.
