应用于嵌入式平台的实时红外行人检测方法

现有基于深度学习的远红外图像行人检测方法对计算力要求高,需要高功耗GPU计算平台,应用于嵌入式平台时,无法满足实时性和准确率需求。针对该问题,本文提出了一种新型实时红外行人检测方法,该方法使用MobileNet作为YOLOv3模型中的基础网络,辅助预测网络层以深度可分离卷积替换标准卷积,将模型改进为轻量红外行人检测模型。基于新方法构建的模型采用CVC红外行人训练集离线训练,并部署于嵌入式平台,实现红外行人在线实时检测。实验结果表明,与改进前方法相比,模型大小为65 M,约为YOLOv3的27%,新模型在基本保证原有准确率的同时,大幅降低了计算量,在同一平台下的检测速度从3FPS提升到了11FPS,可满足大部分嵌入式系统对行人检测的实时性需求。

Real-time infrared pedestrian detection method applied to embedded platform

The far-infrared image pedestrian detection method based on deep learning has high computational power requirements and requires a high-power GPU computing platform. When applied to an embedded platform,it cannot meet the real-time and accuracy requirements. Aiming at this problem,a new real-time infrared pedestrian detection method is proposed in the paper,which uses MobileNet as the basic network in the YOLOv3 model,and assists the prediction network layer to replace the standard convolutions with depth-wise separable convolutions,and improves the model to lightweight infrared pedestrian detection model. The model is built based on the new method uses CVC infrared pedestrian set offline training,and is deployed on the embedded platform to realize infrared pedestrian online real-time detection. The experimental results show that compared with the pre-improvement method,the model size is 65 M,which is about 27 % of YOLOv3. The new model basically reduces the calculation amount while ensuring the original accuracy,and the detection speed under the same platform is from 3FPS upgraded to 11FPS to meet the real-time needs of pedestrian detection in most embedded systems.