基于改进YOLOX的行人检测与跟踪算法

Pedestrian Detection and Tracking Algorithm Based on Improved YOLOX

摘要: 行人多目标跟踪(MOT)在智能交通、自动驾驶及公共安全监控领域应用广泛。针对现有跟踪系统存在的小目标检测精度低、密集场景易漏检、遮挡条件下身份频繁切换、轻量化设计与检测性能难以兼顾等问题，本文提出一种基于YOLOX与ByteTrack的改进型行人检测跟踪框架SSA-YOLOX。该框架引入SPHead小目标检测分支，强化浅层高分辨率特征提取，提升远距离行人与密集目标的感知能力；在特征融合阶段嵌入无参SimAM注意力机制，增强特征判别性并抑制背景干扰；采用轻量化ADown下采样模块，降低计算复杂度且保留关键特征信息；将优化后的检测器与ByteTrack集成，构建完整的多目标跟踪系统。在MOT17数据集上的实验结果表明，SSA-YOLOX以8.92 M参数量实现94.1%的AP@0.5，较原始YOLOX提升2个百分点；跟踪指标达到84.5%的MOTA与80.4%的IDF1，身份切换次数显著降低。所提方法在检测精度、跟踪鲁棒性与模型轻量化之间实现良好平衡，可满足实时行人跟踪的工程应用需求。

Abstract: Pedestrian Multi-Object Tracking (MOT) is widely applied in intelligent transportation, autonomous driving and public security monitoring. To address the problems of low small-object detection accuracy, frequent missed detections in dense scenes, identity switches under occlusion and the trade-off between lightweight design and detection performance, this paper proposes SSA-YOLOX, an improved pedestrian detection and tracking framework based on YOLOX and ByteTrack. The framework introduces the SPHead small-object detection branch to strengthen shallow high-resolution feature extraction and improve the perception of distant pedestrians and dense targets; embeds the parameter-free SimAM attention mechanism into the feature fusion stage to enhance feature discrimination and suppress background interference; adopts the lightweight ADown downsampling module to reduce computational complexity while preserving key feature information; and integrates the optimized detector with ByteTrack to construct a complete multi-object tracking system. Experiments on the MOT17 dataset show that SSA-YOLOX achieves 94.1% AP@0.5 with 8.92 M parameters, which is 2% higher than the original YOLOX. The tracking metrics reach 84.5% MOTA and 80.4% IDF1, and the number of identity switches is significantly reduced. The proposed method achieves a favorable balance among detection accuracy, tracking robustness and model lightweight, and can meet the engineering application requirements of real-time pedestrian tracking.