Scholarlink

<p>Multi-Object Tracking (MOT) is unreplaceable in computer vision for its wide applications. Though Graph Neural Networks (GNNs) are used in existing methods, ways to build graphs and use interactive features generated by GNNs are still open problems. We design an MOT model based on weighted Graph Convolutional Neural Networks (GCNNs). Specifically, we first use the Re-Identification module to extract the appearance features of detections. For tracklet features, we use both the average appearance features of objects within the tracklet and the features of the last object in the tracklet. Then, different from GCNNMatch, we calculate the initial-edge feature by adding cosine similarity constraint and multilayer perceptrons to include more features. For feature updating, we design GCNNs with two node-updating layers and two edge-updating layers. Finally, a fusion module fuses multi-features including interactive cosine similarity, updated edge features, cosine similarity between tracklet appearance and detection ones to get affinity scores which are used to calculate matching results by a designed two-step association algorithm. Our tracker achieves MOTA, IDF1, and HOTA of 79.89%, 78.38% and 64.07% on MOT16; 80.27%, 78.17%, and 63.84% on MOT17; and 77.63%, 76.33% and 62.56% on MOT20 demonstrating its robust data association ability.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multiple object tracking using weighted graph convolutional neural networks

  • 期刊论文

  • 发表日期：

    2025年7月15日

查看全文
  • Yubo Zhang,
  • Liying Zheng,
  • Qingming Huang

摘要

Multi-Object Tracking (MOT) is unreplaceable in computer vision for its wide applications. Though Graph Neural Networks (GNNs) are used in existing methods, ways to build graphs and use interactive features generated by GNNs are still open problems. We design an MOT model based on weighted Graph Convolutional Neural Networks (GCNNs). Specifically, we first use the Re-Identification module to extract the appearance features of detections. For tracklet features, we use both the average appearance features of objects within the tracklet and the features of the last object in the tracklet. Then, different from GCNNMatch, we calculate the initial-edge feature by adding cosine similarity constraint and multilayer perceptrons to include more features. For feature updating, we design GCNNs with two node-updating layers and two edge-updating layers. Finally, a fusion module fuses multi-features including interactive cosine similarity, updated edge features, cosine similarity between tracklet appearance and detection ones to get affinity scores which are used to calculate matching results by a designed two-step association algorithm. Our tracker achieves MOTA, IDF1, and HOTA of 79.89%, 78.38% and 64.07% on MOT16; 80.27%, 78.17%, and 63.84% on MOT17; and 77.63%, 76.33% and 62.56% on MOT20 demonstrating its robust data association ability.