摘要
智能交通的中重要一点就是对行人的检测跟踪用以规避行人实现自动驾驶。在神经网络运用的目标检测之前,常见的行人检测方法有梯度直方图特征(Histogram of oriented gradient,HOG)与支持向量机(Support vector machine,SVM)相结合的方法,但是此方法的弊端较为突出,在检测准确率远远达不到目前主流的几种深度学习算法检测,其应用场景受到很大的限制。而目前在深度学习中,行人检测应用比较广泛的模型有Faster R-CNN、YOLOv3等,而论文将采用改进Cascada R-CNN模型,其比Faster R-CNN具有更好的抗干扰能力,在昏暗、光线不均匀等条件下具有较好的行人检测效果,同时使其能过更好的识别小样本行人,实验在INRIA数据集中完成训练与并在自制的测试集检测,取得了不错的效果。
One of the important points of intelligent transportation is to detect and track pedestrians to avoid pedestrians and realize automatic driving. Before the application of neural network in target detection,the common pedestrian detection methods are histogram of oriented(HOG)and support vector machine(SVM)Machine,(SVM),but the disadvantages of this method is more prominent,in the detection accuracy rate is far from reaching several mainstream deep learning algorithm detection,its application scenarios are greatly restricted. At present,in the deep learning,pedestrian detection models are more widely used,such as Fast R-CNN,YOLOv3,etc. In this paper,the improved Cascade R-CNN model is used,which is better than Faster R-CNN has better anti-interference ability,and has better pedestrian detection effect under the conditions of dim and uneven light. At the same time,it can better identify small samples of pedestrians. The experiment completes training in INRIA data set and detectes in self-made test set,and achieves good results.
作者
贾叙文
刘庆华
刘东华
李杨
黄凯枫
JIA Xuwen;LIU Qinghua;LIU Donghua;LI Yang;HUANG Kaifeng(Jiangsu University of Science and Technology,Zhenjiang 212003)
出处
《计算机与数字工程》
2022年第8期1716-1719,共4页
Computer & Digital Engineering
