Abstract

Deep learning-based perception models are crucial for enabling software-defined vehicles (SDVs) in intelligent transportation systems. However, these models are computationally intensive, demanding substantial processing power and memory resources that often exceed the limits of resource-constrained vehicles. This leads to issues like low throughput, high latency, and excessive GPU and memory usage—challenges that make real-time applications difficult to achieve in SDVs. These limitations affect the model’s responsiveness, efficiency, and overall reliability in dynamic, real-world environments. To address these issues, this talk focuses on our research for enhancing model performance by designing optimized workflows that incorporate pruning and quantization techniques across various computational environments. By leveraging frameworks like PyTorch, ONNX, ONNX Runtime, and TensorRT, we aim to improve efficiency and make perception models more suitable for real-time applications in SDVs. Our findings indicate that a Torch-ONNX-TensorRT workflow, optimized with FP16 precision and group pruning, is the optimal approach for maximizing inference performance, making it well-suited for the demands of resource-constrained SDVs.

Bio

My name is Sumaiya. I received my B.Sc.in Engineering Degree from Khulna University of Engineering and Technology(KUET), Bangladesh in 2020. Currently, I’m a third-year PhD student working towards my PhD degree in Computer Science at Wayne State University, under the supervision of Dr. Zheng Dong. My research interests include autonomous driving perception systems, real-time systems, deep learning, edge-assisted machine learning, and vision systems. I have been working on the acceleration and efficiency of perception systems in edge-assisted autonomous driving. In my second year, I successfully published my initial project at the IEEE MOST 2024 Conference. Building on this work, an extension was later accepted for publication in the IEEE Internet of Things Journal, one of the leading journals in Computing Systems.