卷积神经网络SSD的道路目标检测

Road Target Detection by Convolutional Neural Network SSD

针对在传统的道路目标识别中,需要进行手工提取特征,模型的泛化能力差.使用深度学习的技术,提出了使用深度卷积神经网络(SSD)解决道路目标问题.该方法首先对图像特征进行自动提取,在基础网络后添加不同尺寸的特征图,然后对多尺寸的特征图做卷积滤波,得到目标坐标值和目标的类别.实验中,在SSD模型中增加了特征图的检测层数,增大原图像尺寸,调试相应的参数,经过多次迭代,最终得到目标模型.实验采用行车记录仪采集的图像,在图像中标定出车辆、行人和骑行的人三类,实验表明,检测目标尺寸越小,检测难度越大,检测效果越差,SSD模型对目标检测的平均准确率均值提高了0.082.提出的道路目标检测方法与传统目标识别算法相比,省去了手工特征提取,减少了工作量,提高了模型的泛化能力.

In traditional road target recognition,it was necessary to manually extract features from different environments,and the generalization ability was poor.Using the technology of deep learning,this paper presents a method to solve the road target problem by using the deep convolution neural network(SSD).The method is based on the idea of convolution neural network.The method does automatic extraction on the image features first,and add different size feature maps after the basic network,and then do convolution filtering on the multi dimension feature maps to get the target coordinate value and the target category.The experiment adds the detection layers of the feature maps in the SSD model,increase the size of the original image and the corresponding parameters are optimized and adjusted.The target model is finally obtained after many iterations.In the experiment,we use the image recorded on the tachograph,and identify three kinds of vehicles,pedestrians and riding people in the images.Experiments show that the smaller the detection target size is,the more difficult the detection is,the worse the detection effect is,and improve the recognition of mean average precision of SSD model by 0.082.Compared with the traditional target recognition algorithm,the method of road target detection proposed in this paper omits the process of extracting the artificial features,reduces the workload and improves the generalization ability of the model.