基于嵌入式平台的车前红外行人检测方法研究

Research on Infrared Pedestrian Detection Method in Front of Vehicle Based on Embedded Platform

针对当前目标检测方法普遍需要高功耗GPU计算平台、易受光照条件影响的问题,提出2种基于嵌入式平台的车前红外行人检测方法:将训练好的YOLOv4-tiny模型使用英伟达开源推理加速库TensorRT进行优化,部署于嵌入式平台;以YOLOv4-tiny模型作为算法的基本架构,结合视觉注意力机制和空间金字塔池化思想,同时增加1个YOLO层,提出YOLOv4-tiny+3L+SPP+CBAM网络模型。将2种方法在FLIR数据集上进行训练与测试,并在Jetson TX2嵌入式平台上进行试验,试验结果表明:相较于原始网络YOLOv4-tiny,所提出的第1种方法平均准确率降低0.54%,推理速度提升86.43%(帧速率达26.1帧/s);提出的第2种方法平均准确率提升16.21%,推理速度降低22.86%(帧速率达10.8帧/s)。2种方法均可兼顾准确率和实时性,能够满足车前红外行人检测的需要。

For the problem that current target detection methods generally require a high-power consumption GPU computing platform and are easily affected by lighting conditions,this paper proposed 2 infrared pedestrian detection methods in front of vehicles based on embedded platform:the trained YOLOv4-tiny model was optimized using NVIDIA’s open source inference acceleration library TensorRT and deployed on the embedded platform;the YOLOv4-tiny model was used as the basic architecture of the algorithm,which was combined with the visual attention mechanism and the spatial pyramid pooling idea,and a YOLO layer was added at the same time,a YOLOv4-tiny+3L+SPP+CBAM network model was proposed.The 2 methods were trained and tested on the FLIR dataset,and tested on the Jetson TX2 embedded platform.The test results show that:compared with the original network YOLOv4-tiny,the average accuracy of the first method is reduced by 0.54%,and the inference speed is increased by 86.43%(frame rate up to 26.1 frame/s),the average accuracy of the second method is improved by 16.21%,and the inference speed is reduced by 22.86%(frame rate up to 10.8 frame/s).Both methods can take into account the accuracy and real-time performance,and meet the needs of infrared pedestrian detection in front of vehicle.