This thesis focuses on developing lightweight semantic segmentation models tailored for resource-constrained applications, effectively balancing accuracy and computational efficiency. It introduces several novel concepts, including knowledge sharing, dense bottleneck, and feature re-usability, which enhance the feature hierarchy by capturing fine-grained details, long-range dependencies, and diverse geometrical objects within the scene. To achieve precise object localization and improved semantic representations in real-time environments, the thesis introduces multi-stage feature aggregation, feature scaling, and hybrid-path attention methods.