Realtime Multiclass Semantic Segmentation of Unstructured Indian Urban Road Scenes using SegNet

This paper aims to develop a real-time multi-class semantic segmentation system to detect roads by modifying the SegNet architecture, essentially incorporating a SoftMax output layer. On this basis, the system will utilize the Indian Driving Dataset (IDD) to classify each of the pixels in input images into 20 different classes: roads, vehicles, pedestrians, and other elements of the urban environment. Trained on a dataset of 7000 images, the efficient encoder-decoder structure of the SegNet model will thus enable it to learn the complex features and variance in kinds of scenes seen in the urban roads. The softmax activation function implemented at the output layer helps the model generate probability distributions for each pixel to all classes, thus yielding more accurate multi-class segmentations. Then, the model will be deployed in real-time on the Jetson Nano; such a board can be used in such applications as autonomous driving and traffic management. Furthermore, the achievements obtained are considered to help advance into more profound, clever, and smart intelligent transportation systems based on a sound solution for real-time road object detection in diverse environments.