Abstract:The advancement of autonomous driving systems hinges on the ability to achieve low-latency and high-accuracy perception. To address this critical need, this paper introduces Dynamic Routering Network (DyRoNet), a low-rank enhanced dynamic routing framework designed for streaming perception in autonomous driving systems. DyRoNet integrates a suite of pre-trained branch networks, each meticulously fine-tuned to function under distinct environmental conditions. At its core, the framework offers a speed router module, developed to assess and route input data to the most suitable branch for processing. This approach not only addresses the inherent limitations of conventional models in adapting to diverse driving conditions but also ensures the balance between performance and efficiency. Extensive experimental evaluations demonstrating the adaptability of DyRoNet to diverse branch selection strategies, resulting in significant performance enhancements across different scenarios. This work not only establishes a new benchmark for streaming perception but also provides valuable engineering insights for future work.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to achieve low - latency and high - precision perception in autonomous driving systems. Although traditional object detection algorithms perform well in various scenarios, they often face latency challenges due to inherent computational latency, which can lead to significant differences between the predicted object positions and the actual positions, thus affecting the decision - making process of the autonomous driving system. To solve these problems, the paper introduces the Dynamic Routing Network (DyRoNet), a low - rank enhanced dynamic routing framework specifically designed for flow perception in autonomous driving systems. ### Main Problems 1. **Diverse Scenario Distributions**: The autonomous driving environment is complex and changeable, and a single perception model has difficulty meeting the requirements of various scenarios. It is necessary to customize perception algorithms to adapt to specific environmental conditions while ensuring that these models can work together. 2. **Balance between Performance and Efficiency**: Although large - scale models may be more accurate, they may increase latency; although small - scale models have a faster inference speed, their accuracy is lower. How to balance performance and efficiency in different driving scenarios is a challenge. ### Solutions DyRoNet solves the above problems in the following ways: - **Dynamic Routing Mechanism**: DyRoNet integrates multiple pre - trained branch networks, and each branch is carefully fine - tuned to operate optimally under different environmental conditions. The core module is the speed router, which evaluates the input data and routes it to the most appropriate branch for processing. - **Low - Rank Adapter**: To reduce the risk of over - fitting, DyRoNet introduces the Low - Rank Adapter (LoRA), allowing for specialized adjustments to specific sub - domains while maintaining the generality of the model. - **Multi - Branch Structure**: DyRoNet adopts a multi - branch structure, and each branch can process the current frame and historical frames, thereby better understanding temporal dynamics and achieving low - latency and high - precision perception. ### Experimental Verification The paper evaluates the adaptability and performance of DyRoNet through extensive experiments. The experimental results show that DyRoNet can adapt to different branch selection strategies and significantly improve performance in multiple scenarios. Specifically: - **Flow - Perception Loss**: Each branch is fine - tuned using its original loss function to ensure effectiveness. The router network selects the optimal branch according to efficiency supervision. - **Effective and Efficient (E2) Loss**: By comprehensively considering hardware differences and inference time costs, the branch selection is optimized to ensure that the model can maintain the best performance under different computing resources and environmental conditions. ### Summary Through dynamic routing and low - rank adapters, DyRoNet effectively solves the challenges of low - latency and high - precision perception in autonomous driving systems, providing valuable engineering insights for future flow - perception research.

DyRoNet: Dynamic Routing and Low-Rank Adapters for Autonomous Driving Streaming Perception

Research and Implementation of Networked Cloud-Controlled Autonomous Driving System

Enabling Robust DRL-Driven Networking Systems Via Teacher-Student Learning

Deep learning based adaptive Ryu controller model for quality of experience issues in multimedia streaming for software defined vehicular networks

Neural Adaptive Transport Framework for Internet-scale Interactive Media Streaming Services

An Integrated DQN and RF Packet Routing Framework for the V2X Network

RoadRunner -- Learning Traversability Estimation for Autonomous Off-road Driving

RLSS: A Reinforcement Learning Scheme for HD Map Data Source Selection in Vehicular NDN

(Pr)-R-3: Realizing Robust Routing For Vanet Using Trajectory Prediction And Crossroad Recognition

StreamYOLO: Real-time Object Detection for Streaming Perception

Deep-Reinforcement-Learning-Based Intelligent Routing Strategy for FANETs

Enabling High-Throughput Routing for LEO Satellite Broadband Networks: A Flow-Centric Deep Reinforcement Learning Approach

G-Routing: Graph Neural Networks-Based Flexible Online Routing

An NN-Based Double Parallel Longitudinal and Lateral Driving Strategy for Self-Driving Transport Vehicles in Structured Road Scenarios

RHRA-DRL: RSU-Assisted Hybrid Road-Aware Routing Using Distributed Reinforcement Learning in Internet of Vehicles

DRRNets: Dynamic Recurrent Routing Via Low-Rank Regularization in Recurrent Neural Networks.

Self-Organized Routing for Autonomous Vehicles Via Deep Reinforcement Learning

Hierarchical framework integrating rapidly-exploring random tree with deep reinforcement learning for autonomous vehicle

Deep Reinforcement Learning for Router Selection in Network with Heavy Traffic

ReasonNet: End-to-End Driving with Temporal and Global Reasoning

Distributed Routing Controller for Large-scale Live Video Streams in Real-Time Networks.