基于改进CornerNet-Lite的林区行人检测算法

Forest-pedestrian detection algorithm based on improved CornerNet-Lite

为做好森林管护工作,减少人力和物力的消耗,利用快速发展的行人检测算法可对林区行人进行检测和甄别,但林区行人检测与传统行人检测有很多不同,如林区行人被树木遮挡、着装与背景色调相似等特征,导致漏检现象严重。为此,针对林区行人的特点,笔者提出了一种实时检测林区行人的算法CornerNet-P,将林区行人的位置预测简化成为2个关键点的预测。以CornerNet-Lite为基础,改进算法的损失函数,预测2组热力图来检测林区行人的角点位置,预测嵌入层损失以匹配同一行人的角点,预测偏置层损失来减少尺度变化过程中的精度损失,并获得最终的边界框;然后提取COCO2014数据集中的行人数据并随机分为训练集和测试集两部分,使用训练集分别训练该算法与YOLOv4算法中的参数,使用测试集和真实的林区行人图像对算法的检测精度和检测速度进行检验。试验结果表明,CornerNet-P算法相比YOLOv4算法平均检测精度提高了1.7%,检测速度提高了5.1%,并可以较好地检测真实林区行人图像。CornerNet-P算法可以实现林区的行人检测,具有较快的检测速度和较满意的精度。

Good forest management and protection is a necessary condition for reducing forest illegal cutting and improving forest development environment.In recent years,with the improvement of computer hardware and the rapid development of deep learning technology,the pedestrian detection algorithm can be used to manage and protect the forest,reduce the consumption of labor and material resources,and detect trespassing pedestrians.The detection of pedestrians immediately reported to the forestry management department,personnel screening and the case of illegal logging in a timely manner.However,due to the lush vegetation in the forest area,it is easy for different degrees of occlusion to occur when collecting images,and the characteristics of pedestrians are not complete.Moreover,most illegal loggers wear camouflage clothing,which is similar to the color of the background,so the detection of pedestrians in the forest area is difficult,and the existing detection algorithm needs to be optimized.A CornerNet-P algorithm for real-time detection of forest pedestrians was proposed to simplify the prediction of the location of forest pedestrians into two key prediction points.Using CornerNet-Lite,this study improved the loss function of the algorithm and predicted two sets of heatmaps to detect the corner position of the forest pedestrian.Then the algorithm predicted the embedded layer loss to match the corner of the same pedestrian and predicted the offset layer loss to reduce the precision loss in the process of scale change,and then the final anchor box was obtained.Then extracted pedestrian data from the COCO2014 dataset and randomly divided it into two parts:a training set and a validation set.The training set was used to train the parameters of the algorithm and YOLOv4 algorithm respectively,and the validation set and real forest pedestrian images were used to test the detection accuracy and speed of the algorithm.The experimental results showed that the CornerNet-P algorithm improved the detection accuracy by 1.7%and the detection speed by 5.1%compared with that of YOLOv4.In general,the CornerNet-P algorithm could achieve a faster detection speed and satisfactory precision for the detection of pedestrians in the forest region.