|
In recent years, researches in the field of autonomous driving are in full swing. Among them, the road detection technology is the key. This technology uses sensors to analyze road forward derection, traffic signs, road lines, pedestrains\' status and other information in real-time. Because the road environment is very complex, the road detection algorithm must be robust against illumination changes, different weather conditions etc. In real life, the road environment changes little, and the daily driving routes are mostly in the same section. Based on these assumptions, this paper propased a way to tranform the road image which are under poor environmental conditions into images that are taken in the same place but are under good environmental conditions, and then to carry out road detection in subsequent mudules. In order to verify the idea, a road image database was established firstly, and the system was completed base on the classical CNN network. The expected results were obtained. In order to improve system\'s accuracy, this paper fine tuned the parameters of the neural network in the classical CNN. Afterwards, the transformation was learned on the basis of the CNN features and then the CNN features were projected into the domain-invariant feature space which was immune to drastic weather or illumination changes. Finally, the experiments have proved the practicality and the validity of the system. |
|
Keywords:deep learning; scene retrieval; road detection; re-identification; neural network |
|