Publication

Abstract : The need for advanced computer vision systems specifically, semantic segmentation models—has increased as autonomous vehicles become more and more integrated into contemporary transportation networks. These systems are necessary for accurate scene perception. However, implementing these models on resource-constrained edge devices poses significant difficulties, including issues with the requirement for high real-time processing speed, high memory usage, and computational inefficiencies. This research work investigates the application of the model quantization approach to address the pressing demand for effective semantic segmentation models optimised for deployment on autonomous cars as a means of bridging the gap between the constraints of edge devices and the complexities of autonomous driving in response to these obstacles. Quantization techniques play a pivotal role in optimizing the inference speed of deep neural networks without compromising their performance. In this research, we explore the impact of employing 8-bit integer quantization on inference time of the state-of-the-art fast semantic segmentation model. Our findings demonstrate a remarkable reduction in inference time, decreasing from 10.4 milliseconds to 3.7 milliseconds. Our quantization module enhances computing speed and power efficiency in edge devices by reducing the weight parameter size, building upon the two-branch approaches for semantic segmentation that are currently in use. The network yields an pixel accuracy of '94.9%' and mean intersection over union of ‘58.31%’ on Cityscapes dataset.