OPTIMIZACIJA MODELA DUBOKOG UČENJA PRIMJENOM KVANTIZACIJE U REAL-TIME SISTEMIMA AUTONOMNIH VOZILA / OPTIMIZATION OF DEEP LEARNING MODELS USING QUANTIZATION IN REAL-TIME SYSTEMS FOR AUTONOMOUS VEHICLES

Abstract

The development of autonomous vehicles depends on the ability of embedded systems to process large volumes of sensor data in real time, collected through cameras, LiDAR, and radar. Deep learning models are used for tasks such as object detection, traffic sign recognition, and lane detection, but their complexity poses challenges for deployment on resource-constrained platforms. Model quantization is an effective optimization technique that reduces the numerical precision of neural network weights and activations. By replacing 32-bit representations with lower precision formats (16-bit or 8-bit), memory usage is reduced and inference speed is improved, which is essential for real-time systems. In addition, techniques such as pruning and knowledge distillation are used to reduce model complexity while maintaining performance. These methods can be combined to achieve an optimal balance between accuracy, speed, and hardware constraints. This paper analyzes the role of quantized models in autonomous vehicles, highlighting their advantages and limitations, with a focus on implementation on GPUs, TPUs, and AI accelerators and their impact on system performance.