Abstract: To address the issues of false detection and missed detection caused by severe target occlusion and the lack of consideration for the relationships between multiple views in existing multi-view pedestrian detection algorithms, this paper proposes an improved multi-view pedestrian detection method based on the MVDetr algorithm. First, during the feature extraction phase, a View Enhancement Module (VEM) is designed to enhance important views by focusing on the relationships between different views. Second, in the process of introducing multi-view information into a single view, an Efficient Multi-scale Attention module (EMA) is added to establish short-distance and long-distance dependencies, thereby improving detection accuracy. Finally, based on the original Shadow Transformer module in the baseline algorithm, a new multi-view information processing module called Efficient Shadow Transformer (EST) is designed to reduce the use of redundant information in multiple views while maintaining detection performance. Experimental results show that the improved algorithm enhances the main detection metric MODA by 1.8 percentage points, MODP by 0.6 percentage points and Recall by 1.8 percentage points on the Wildtrack dataset compared to the original MVDetr algorithm, demonstrating its effectiveness for multi-view pedestrian detection tasks.

Key words: multi-view, pedestrian detection, MVDetr, attention mechanism, feature enhancement

摘要： 针对现有的多视角行人检测算法中因目标遮挡严重、未关注多视角之间关系而导致错检、漏检等问题，提出一种基于MVDetr算法改进的多视角行人检测方法。首先，在特征提取阶段，设计了一个视角特征增强模块VEM(View Enhancement Module)，通过关注不同视角之间的关系实现对重要视角的增强；其次，在将多视角信息引入单视角的过程中，加入高效多尺度注意力模块EMA(Efficient Multi-Scale Attention)，建立短距离和长距离依赖关系，提高检测精度。最后，在原始基线算法Shadow Transformer模块的基础上，设计了一种新的多视角信息处理模块EST(Efficient Shadow Transformer)，能够在保持检测效果的基础上，减少多视角中冗余信息的使用。实验结果表明，改进后的算法在Wildtrack数据集上与原始MVDetr算法相比，主要检测指标MODA提升了1.8个百分点，MODP提升了0.6个百分点，Recall提升了1.8个百分点，能够很好地应用于多视角行人检测任务。

关键词: 多视角, 行人检测, MVDetr, 注意力机制, 特征增强