Embedded Files
This website provides supplementary materials for the paper "When Autonomous Vehicle Meets V2X Cooperative Perception: How Far Are We?".
Implementation Details
Experimental Runtime Environment
The offline evaluation parts are performed on a Linux Mint 21.3 server with 4 GeForce RTX 2080 Ti GPUs, two 20-core processors with 2.0 GHz, and 125GB RAM. And the online evaluation parts are performed on an Ubuntu 20.04.6 LTS server with one Nvidia GeForce RTX 4070 Ti GPU, one 8-core processor with 3.3 GHz, and 32GB RAM. Overall, the online evaluation simulated an effective runtime of approximately 110 hours for the ADS equipped with cooperative perception.
Supplementary Experimental Results and Analysis
To facilitate clearer interpretation of the findings, we present the experimental results in graphical form and provide a corresponding analysis.
RQ1
(1) Discussion of the results in the paper: "Moreover, the total number of CCME and CCLE errors in camera-based cooperative perception is substantially higher than in LiDAR-based cooperative perception, leading to a pronounced decline in the AP of the camera-based configuration."
Proportion of in CCME and CCLE under heterogeneous sensor configurations.
(2) Discussion of the results in the paper: "Additionally, we observe that the number of cooperative localization errors is, on average, 122.0% and 38.8% greater than the number of cooperative missing errors and additional detection errors, respectively, across all configurations." Here, cooperative localization errors includes LCLE and CCLE, cooperative missing errors includes LCME and CCME, and additional detection errors includes LADE and CADE.
Comparison of cooperative perception errors across all configurations in each frame.
(3) Discussion of the results in the paper: "Furthermore, in most configurations, the number of miscorrected cooperative errors exceeds that of misleading cooperative errors. This indicates that the majority of errors arise from the cooperative perception system's failure to correct inaccurate or incomplete information initially predicted by the ego vehicle."
The number of misleading cooperative error and miscorrected cooperative error in each frame.
RQ2
(4) As shown in the table below, the AP of V2V cooperative perception is 3.4% and 4.8% higher than that of V2I under intermediate and late fusion schemes, respectively. In contrast, V2I achieves 3.4% higher AP than V2V under the early fusion scheme.
Cooperative error patterns analysis of the system at varying distances under V2V and V2I cooperation.
(5) For intermediate fusion, V2V cooperation results in a higher number of CCLE, while V2I cooperation exhibits a greater number of CCME.
The number of CCME and CCLE for different cooperative modes in each frame.
RQ3
(6) Discussion of the results in the paper: "The results show that the frequency of cooperative perception errors increases as the time of the violation approaches, with particularly notable increases in CCME and CCLE."
Distribution of cooperative perception errors over the entire time period from ADS initiation to violation occurrence.
Page updated
Google Sites
Report abuse