轻量化深度学习模型在行人检测上的应用

Application of Lightweight Deep Learning Model in Pedestrian Detection

摘要: 行人检测和跟踪在目标跟踪领域至关重要，广泛应用于辅助驾驶、安全监测和其他行人分析。在多目标跟踪中，面临多种挑战，因此需要设计实时性和高精度的算法。本研究提出了一种新的行人跟踪模型。在行人特征建模阶段，采用Yolov4-tiny网络模型和COCO数据集预训练权重参数，经过迁移学习到MOT数据集。为了解决目标微小部分的变形和遮挡问题，引入了一种深度分类跟踪器，结合了MeanShift滤波器和卡尔曼滤波器。通过反投影图像和物体轮廓与卡尔曼线性观测模型相融合，实现了目标预测。实验结果表明，该模型能够在复杂环境中长时间跟踪目标，具有良好的跟踪效果，多目标跟踪精度为57.6%，目标定位精度为82.1%。

Abstract: Pedestrian detection and tracking are crucial in the field of target tracking and are widely used in assisted driving, safety monitoring, and other pedestrian analysis. In multi-target tracking, various challenges are faced; therefore, it is necessary to design real-time and high-precision algorithms. This study proposes a new pedestrian tracking model. In the pedestrian feature modeling stage, the Yolov4 tiny network model and COCO dataset were used to pretrain weight parameters, which were then transferred and learned to the MOT dataset. In order to solve the problem of deformation and occlusion of small parts of the target, a deep classification tracker is introduced, which combines the MeanShift filter and the Kalman filter. By integrating back projection images and object contours with Kalman linear observation models, target prediction has been achieved. The experimental results show that the model can track targets for a long time in complex environments and has good tracking performance. The multi-target tracking accuracy is 57.6%, and the target positioning accuracy is 82.1%.