FURI | Spring 2025

Compression of Deep Neural Networks for Deployment on Edge Devices

Large deep neural networks (DNNs) are essential for autonomous driving tasks such as object detection, lane tracking, and collision warnings. However, they pose significant deployment challenges on resource-constrained edge devices. This research study investigates how compression algorithms, specifically quantization, can reduce the size of depth models while retaining at least 90% of the original model’s accuracy while increasing real-time inference speed. The NVIDIA Jetson Nano-based Duckiebot DB21 is the testbed for real-time inference. If successful, inference and decision-making can be much quicker than relying on larger external computational resources.