Multiagent Detection System Based on Spatial Adaptive Feature Aggregation

Hongbo Wang

doi:10.1109/JSYST.2024.3423752

Return

Multiagent Detection System Based on Spatial Adaptive Feature Aggregation

2024-12-01

PRE

DOI：10.1109/JSYST.2024.3423752

Original

Original request for help

Save

Abstract

En 中文

Detection systems based on computer vision play important roles in Large-Scale Multiagent Systems. In particular, it can automatically locate and identify key objects and enhance intelligent collaboration and coordination among multiple agents. However, classification and localization in object detection may produce inconsistent prediction results due to different learning focus. Therefore, we propose a Spatial Decoupling and Boundary Feature Aggregation Network (SDBA-Net) to achieve spatial decoupling and task alignment. SDBA-Net includes a spatially sensitive region-aware module (SSRM) and a boundary feature aggregation module (BFAM). SSRM predicts sensitive regions for each task while minimizing computational cost. BFAM extracts valuable boundary features within sensitive regions and aligns them with corresponding anchors. These two modules are combined to spatially decouple and align the features of two tasks. In addition, a significance dependency complementary module (SDCM) is introduced. It enables SSRM to quickly adjust the sensitive region of the classification task to the significant feature region. Experiments are conducted on a large-scale complex real-world dataset MS COCO (Lin et al., 2014). The results show that SDBA-Net achieves better results than the baselines. Using the ResNet-50 backbone, our method improves the average precision (AP) of the single-stage detector VFNet by 1.0 point (from 41.3 to 42.3). In particular, when using the Res2Net-101-DCN backbone, SDBA-Net achieves an AP of 51.8 on the MS COCO test-dev.

Keywords:

Feature extraction

Task analysis

Location awareness

Accuracy

Poles and towers

Object detection

Natural resources

Complex scenarios

feature aggregation

large scale

multiagent