C-MASS: Combinatorial Mobility-Aware Sensor Scheduling for Collaborative Perception with Second-Order Topology Approximation

TL;DR

This work tackles perception under bandwidth constraints in collaborative perception (CP) by modeling the perception topology with a second-order approximation and maintaining an empirical topology via a replay mechanism. The authors introduce C-MASS, a pull-based, object-oriented CP framework that uses data replay and motion-based prediction to keep an up-to-date topology while scheduling CoVs under a bandwidth budget. To optimize scheduling, they design a hybrid greedy algorithm with a worst-case performance guarantee for the resulting budgeted maximum coverage variant; they further balance exploration and exploitation using an upper-confidence-bound style mechanism that weighs topological uncertainty. Numerical experiments on SUMO-based mobility traces and datasets OPV2V/V2VReal show that C-MASS achieves near-optimal performance, with significant improvements in recall and weighted recall over object-level CP and heuristic baselines, confirming the practical impact for edge-assisted and distributed CP configurations.

Abstract

Collaborative Perception (CP) has been a promising solution to address occlusions in the traffic environment by sharing sensor data among collaborative vehicles (CoV) via vehicle-to-everything (V2X) network. With limited wireless bandwidth, CP necessitates task-oriented and receiver-aware sensor scheduling to prioritize important and complementary sensor data. However, due to vehicular mobility, it is challenging and costly to obtain the up-to-date perception topology, i.e., whether a combination of CoVs can jointly detect an object. In this paper, we propose a combinatorial mobility-aware sensor scheduling (C-MASS) framework for CP with minimal communication overhead. Specifically, detections are replayed with sensor data from individual CoVs and pairs of CoVs to maintain an empirical perception topology up to the second order, which approximately represents the complete perception topology. A hybrid greedy algorithm is then proposed to solve a variant of the budgeted maximum coverage problem with a worst-case performance guarantee. The C-MASS scheduling algorithm adapts the greedy algorithm by incorporating the topological uncertainty and the unexplored time of CoVs to balance exploration and exploitation, addressing the mobility challenge. Extensive numerical experiments demonstrate the near-optimality of the proposed C-MASS framework in both edge-assisted and distributed CP configurations. The weighted recall improvements over object-level CP are 5.8% and 4.2%, respectively. Compared to distance-based and area-based greedy heuristics, the gaps to the offline optimal solutions are reduced by up to 75% and 71%, respectively.

Figures (12)

• Figure 1: Illustrations of the two configurations of CP. (a) Edge-assisted CP with V2I/I2V communication. (b) Distributed CP with V2V communication.
• Figure 2: The LiDAR point clouds and detection results of a frame in V2V4Real. (a) The ego vehicle only. (b) The auxiliary vehicle only. (c) Feature-level CP. Red boxes: ground truth objects. Green boxes: objects detected by CoBEVT.
• Figure 3: 3D surface graphs of conditional missed detection probabilities given the number of scanned points at both perspectives. (a) Statistics from V2V4Real. (b) Fitted model for V2V4Real.
• Figure 4: The workflow of the C-MASS framework.
• Figure 5: Illustration of the replay module.

Theorems & Definitions (10)

• Definition 1
• Definition 2
• Lemma 1
• proof
• Lemma 2
• proof
• Lemma 3
• proof
• Theorem 1
• proof