针对红外图像纹理弱及多目标遮挡导致跟踪精度低的问题,构建了基于改进YOLOv7模型和多目标跟踪算法DeepSort的融合红外目标跟踪模型MSB-YOLOv7-DeepSort.采用SE(squeeze and excitation)通道注意力机制和双向特征金字塔网络提高红外目标的特征提取质量;利用轻量化网络MobileNetV3替换YOLOv7骨干网络,提升融合模型的推理速度.实验结果表明,MSB-YOLOv7-DeepSort模型在跟踪准确度、跟踪精确度、正确目标跟踪比例和帧率等方面均具有较好的性能.
Abstract
To solve the problem of low tracking accuracy caused by weak texture and multi target occlusion,a fusion infrared tracking model MSB-YOLOv7-DeepSort is constructed,based on the improved YOLOv7 and the multi-target tracking algorithm DeepSort.The SE(squeeze and excita-tion)channel attention mechanism and the bidirectional feature pyramid network are utilized to im-prove the quality of feature extraction for the infrared targets.The lightweight network Mobile-NetV3 is used to replace the YOLOv7 backbone network to enhance the inference speed of the fu-sion model.The experimental results indicate that the MSB-YOLOv7-DeepSort model has good per-formance in tracking accuracy and frame rate.
维普
万方数据