Invisible Triggers, Visible Threats! Road-Style Adversarial Creation Attack for Visual 3D Detection in Autonomous Driving

MMEdge is a new framework designed for real-time multimodal inference on resource-constrained edge devices, crucial for applications like autonomous driving and mobile health. The framework addresses the challenges of sensing dynamics and model execution by decomposing the inference process into fine-grained sensing and encoding units. This allows for incremental computation as data arrives, while a lightweight temporal aggregation module captures rich temporal dynamics to maintain accuracy.

The article discusses the growing interest in world models, particularly in light of advancements in multimodal large language models like GPT-4 and video generation models such as Sora. World models serve two main purposes: understanding current world mechanisms and predicting future dynamics. The review categorizes these models and highlights their applications in various fields, including generative games, autonomous driving, robotics, and social simulacra, emphasizing their role in decision-making processes.

The article discusses a novel framework called Learning with Preserving (LwP) designed for Continual Multitask Learning (CMTL) in artificial intelligence systems. CMTL involves models that learn new tasks sequentially without forgetting previously acquired skills, which is crucial in fields like autonomous driving and medical imaging. Traditional methods often struggle due to task-specific feature fragmentation. LwP focuses on maintaining the geometric structure of shared representation spaces, enhancing the model's ability to learn continuously.

The article discusses a novel adaptive LiDAR scanning framework that enhances 3D object detection by utilizing temporal cues from past observations. Traditional LiDAR sensors often perform redundant scans, leading to inefficiencies in data acquisition and power consumption. The proposed method employs a lightweight predictor network to identify regions of interest, significantly reducing unnecessary data collection and improving overall efficiency.

Backdoor attacks pose a significant risk to deep learning systems, especially in critical 3D applications like autonomous driving and robotics. This study introduces a novel one-to-N backdoor framework for 3D vision, utilizing a configurable spherical trigger. The research demonstrates that a single trigger can effectively encode multiple target classes, achieving high attack success rates of up to 100% while preserving accuracy on clean data.

The CATS-V2V dataset introduces a pioneering real-world collection for Vehicle-to-Vehicle (V2V) cooperative perception, aimed at enhancing autonomous driving in complex adverse traffic scenarios. Collected using two time-synchronized vehicles, the dataset encompasses 100 clips featuring 60,000 frames of LiDAR point clouds and 1.26 million multi-view camera images across various weather and lighting conditions. This dataset is expected to significantly benefit the autonomous driving community by providing high-quality data for improved perception capabilities.

The paper titled 'FQ-PETR: Fully Quantized Position Embedding Transformation for Multi-View 3D Object Detection' addresses the challenges of deploying PETR models in autonomous driving due to their high computational costs and memory requirements. It introduces FQ-PETR, a fully quantized framework that aims to enhance efficiency without sacrificing accuracy. Key innovations include a Quantization-Friendly LiDAR-ray Position Embedding and techniques to mitigate accuracy degradation typically associated with quantization methods.

The paper titled 'FQ-PETR: Fully Quantized Position Embedding Transformation for Multi-View 3D Object Detection' addresses the challenges of deploying PETR models in autonomous driving due to their high computational costs and memory requirements. It introduces FQ-PETR, a fully quantized framework that aims to enhance performance while maintaining accuracy. The proposed innovations include a Quantization-Friendly LiDAR-ray Position Embedding and improvements in quantizing non-linear operators, which are critical for effective multi-view 3D detection.