Hybrid Deep Learning Aerial Framework for Road Scene Objects Segmentation and Classification
Aysha Naseer, Ahmad Jalal
Abstract
This research proposes an advanced approach of object segmentation and categorization using aerial image sequences for enhancing intelligent traffic monitoring systems. Traditional approaches are generally time-consuming for calculations and are distorted by variation in gathering data. In an effort to meet these challenges a five-point strategy is suggested. First prior to feeding into the model, input images are augmented and cleaned up for noise. Foreground objects are then extracted using Deep Learning based segmentation with the help of DeepLabV3+model incorporated with VGG encoder. To enhance the performance, the model deployed deep features from VGG and handcrafted features such as Gabor Filters and LBP for better feature description of objects’ texture and depth. All these fused features are fed to identify a CNN classifier which boosts the accuracy of the object classification. The suggested model provides the highest accuracy of $\mathbf{9 7. 8 0 \%}$ and $\mathbf{9 7. 3 0 \%}$ on benchmark VEDAI and VAID datasets respectively, thus outperforming the conventional methodologies. Comparisons with traditional methods prove that the current hybrid approach is uniquely efficient in terms of computational speed as well as providing more accurate results for the application in intelligent traffic monitoring and similar fields.