基于改进YOLOv7-tiny的多尺度运动目标检测算法

Multi-scale Moving Object Detection Algorithm Based on Improved YOLOv7-tiny

针对区域安全防卫时,远距离入侵目标成像像素少,纹理信息不足以及在目标持续接近过程中的大尺度变换导致的模型误检、漏检问题,本文提出一种基于YOLOv7-tiny算法改进的多尺度运动目标检测算法。首先,提出一种新的OBM模块用于特征提取网络,利用多维注意力机制提高网络的特征提取能力;其次,采用改进的AC-BiFPN双向特征融合策略,将进行自适应加权融合后的多层次特征传给ACmix注意力机制,以提升模型对多尺度目标的感知能力;最后,优化模型的损失函数,对预测框与真实框之间的区域加权,减少模型预测偏差。模型在自制的监控行人车辆数据集上进行实验,结果表明,相较于原始的YOLOv7-tiny模型,改进后的YOLOv7-tiny模型改善了远距离监测时行人车辆的误检漏检问题,检测精确度提高了3.96个百分点,平均检测精度(mAP@0.5:0.95)提高了2.22个百分点,在边缘GPU上实时帧率达到32.7 fps,满足实际使用需求。

Aiming at the problem of model misdetection and omission caused by the few pixels of long-distance intrusion object,insufficient texture information as well as large-scale transformations during the continuous approach of the object in regional security defence,an improved multi-scale moving object detection algorithm based on the YOLOv7-tiny algorithm is proposed.Firstly,a new OBM module is proposed for the feature extraction network,using a multi-dimensional attention mechanism to improve the feature extraction capability of the network.Secondly,an improved AC-BiFPN bidirectional feature fusion strategy is used to combine the multi-dimensional adaptive weighted fusion.The scale features are passed to the ACmix attention mechanism to improve the model’s perception of multi-scale objects.Finally,the activation function of the model is optimized to weight the area between the predicted frame and the real frame to reduce the model prediction bias.The model is tested on a self made pedestrian and vehicle data set,and the experimental results show that compared with the original YOLOv7-tiny model,the improved YOLOv7-tiny model addresses the problem of misdetection and omission of pedestrians and vehicles during long distances monitoring,with an increase of 3.96 percentage points in the detection accuracy,an increase of 2.22 percentage points in the average detection accuracy(mAP@0.5:0.95),and the real-time frame rate reaches 32.7 fps on edge GPUs,which meets the practical application requirements.