Improved Decision-Making for Navigation Scenarios

Nowadays, intelligent systems are quite common to assist the driver in its navigation chores. But, if accurate information is not provided to the driver on time then it is of no use and the delay can even lead to disastrous consequences. To improve behavioral realism in real life navigation scenarios we need intelligent and speedy decisions. Such decisions rely on the information being extracted from real-time images. We need to select the best available features/information that can be used as neural network inputs and ultimately predict the next move of the vehicle. Little work has been done on selection of best features of navigation images and research needs to be done in this gray area as its results have direct impact on the classification accuracy and generalized performance of automotive navigation planning. This article proposes a novel approach to find the best possible set of features from real-time navigation images by using Boruta Algorithm and Earth Algorithm so as to improve the prediction power of vehicle to either move or stop in the next course of action. The results obtained were cross-validated by using three classifiers: Logistic Regression, Support Vector Machines (SVM), and K-Nearest Neighbor on the basis of parameters: Classification accuracy and prediction performance. It was identified that the proposed model maintained the classification accuracy and performed with more superiority by getting rid of the irrelevant features of images and thereby reducing the training time as compared to computations done on the basis of original feature set. The prediction speed of the proposed model was found to be much better than the model without feature selection. The accuracy of this novel approach was also found to be improved by few ensembles using Generalized Linear Model (GLM) wrapper.