Cooper: Cooperative Perception for Connected Autonomous Vehicles based on 3D Point Clouds (1905.05265v1)

Abstract: Autonomous vehicles may make wrong decisions due to inaccurate detection and recognition. Therefore, an intelligent vehicle can combine its own data with that of other vehicles to enhance perceptive ability, and thus improve detection accuracy and driving safety. However, multi-vehicle cooperative perception requires the integration of real world scenes and the traffic of raw sensor data exchange far exceeds the bandwidth of existing vehicular networks. To the best our knowledge, we are the first to conduct a study on raw-data level cooperative perception for enhancing the detection ability of self-driving systems. In this work, relying on LiDAR 3D point clouds, we fuse the sensor data collected from different positions and angles of connected vehicles. A point cloud based 3D object detection method is proposed to work on a diversity of aligned point clouds. Experimental results on KITTI and our collected dataset show that the proposed system outperforms perception by extending sensing area, improving detection accuracy and promoting augmented results. Most importantly, we demonstrate it is possible to transmit point clouds data for cooperative perception via existing vehicular network technologies.

• The paper introduces the SPOD method, which significantly enhances detection accuracy in diverse LiDAR configurations through cooperative data fusion.
• It employs an end-to-end deep neural network with voxel-based feature extraction and sparse convolutional networks to process both high- and low-resolution point clouds.
• Experimental evaluations on benchmark datasets validate improved object detection performance and safety, underscoring the benefits of shared vehicular perception.

Cooper: Cooperative Perception for Connected Autonomous Vehicles Using 3D Point Clouds

The paper "Cooper: Cooperative Perception for Connected Autonomous Vehicles based on 3D Point Clouds" by Qi Chen, Sihai Tang, Qing Yang, and Song Fu introduces an innovative approach to improving the perception capabilities of autonomous vehicles through a cooperative network leveraging 3D LiDAR point clouds. The necessity arises from the known limitations in individual vehicular perception systems which can lead to critical object detection failures and subsequent safety compromises.

Overview

The premise of the paper is the enhancement of object detection accuracy through a cooperative perception framework for connected autonomous vehicles (CAV). The researchers propose a method named Sparse Point-cloud Object Detection (SPOD), which operates on low-density point clouds collected via LiDAR sensors. This method addresses the unwilling density variability in data captured by different LiDAR configurations, prominently among 64-beam and 16-beam devices, thus broadening the utility across a diversified vehicular inventory.

Detection Methodology

The SPOD system employs an end-to-end deep neural network capable of effectively processing both high and low-resolution raw LiDAR inputs. This innovation is crucial as existing CNN-based object detection methodologies tend to falter when applied to sparse datasets typical in lower-end LiDAR configurations. The detection architecture leverages voxel-based feature extraction and Sparse Convolutional Neural Networks (SCNN), ensuring efficiency in computational resource usage while maintaining robust object identification performance.

Empirical Evaluation

Experiments conducted using both the KITTI dataset and a custom T${content}$J dataset validate the theoretical underpinnings of the Cooper system. In fragmented evaluative scenarios, the research delineates substantial gains in object detection scores and overall information acquisition, particularly with cooperative perception mechanisms in play. Consistently, through cooperative fusion of point clouds both spatial coverage and object detection accuracy saw measurable improvement. Notably, new object identifiers emerged upon integration of cooperative data over traditional standalone analyses.

Practical and Theoretical Implications

The implication of this paper primarily circles around vehicular safety and reliability amelioration. By expanding the detection horizon through cooperative sensing and data sharing, vehicles can potentially mitigate risks associated with sensor line-of-sight limitations and data sparsity. From a practical networking perspective, the Rosen's transmission and storage optimization method underlines an economically-viable bandwidth utilization, crucial for enabling real-time data sharing in bandwidth-constrained vehicular networks like DSRC.

Theoretically, this paper advances the foundational understanding of multi-agent and sensor fusion systems within autonomous vehicle ecosystems. It challenges existing paradigms centered around sole-vehicle perception autoregulation, advocating instead for a symbiotic data sharing framework.

Future Directions

Future research avenues will likely delve into the challenges of heterogeneity in sensor data across diverse vehicle makes, alongside further optimizations for minimizing data exchange latency and maximizing real-time perception accuracy. Comparative studies involving different adaptive networks (e.g., 5G, V2X) could optimize the cooperative perception systems further.

Moreover, integrating such cooperative systems with other sensory data modalities, such as radar or camera feeds, may refine SPOD and similar algorithms, offering even greater operational fidelity and safety guarantees in convoluted driving environments.

Conclusion

The Cooper system's pioneering approach to cooperative vehicular perception harnesses the strengths of collective intelligence, manifesting notable improvements in detection efficacy and safety for autonomous vehicles. This paper's methodologies and findings not only lay a practical groundwork but also propel future discourse on AI-driven networked vehicular systems, beckoning further exploration and implementation.