基于上下文以及多尺度信息融合的目标检测算法

Object Detection Algorithm Based on Context and Multi-scale Information Fusion

卷积神经网络的快速发展极大地提升了目标检测的性能。针对SqueezeDet算法没有利用多尺度以及上下文信息的问题,文章结合跳过连接(skip connection)和快捷连接(shortcut connection)来汇聚多尺度特征图,利用膨胀卷积(dilated convolution)来扩大卷积感受野以及上下文信息,提出了一种基于上下文的多尺度目标检测模型,提升了整个网络对复杂场景下的目标检测的精度和鲁棒性。该模型融合3种不同分辨率的特征图:将最小以及中间尺寸的特征图通过不同采样率的膨胀卷积聚集上下文信息,然后通过双线性插值的方式将最小特征图的分辨率放大一倍,最大特征图经卷积层降采样之后获得与中间特征图相同的尺寸,与之进行融合,并且使用了快捷连接来连接不同尺寸的特征图,从较大特征图中获取丢失的信息。将该模型在自动驾驶国际公开基准测试数据集KITTI中进行了实验,与SqueezeDet相比,所提算法的准确率提升约5%,同时在GPU中的推断速度可达30 fps。

Recent advances in convolutional neural networks(CNNs)have led to significant improvement in object detection.To solve the problem of missing context and multi-scale information of SqueezeDet algorithm,this paper combines skip connection and shortcut connection to aggregate multi-scale feature maps,and use dilated convolution to expand the convolutional receptive field and context.A context-based multi-scale object detection model was proposed to effectively improve the accuracy and robustness of object detection for complex scenes.This model fuses three different resolution feature maps:the minimum and middle size feature maps gather context through dilated convolution,the minimum size feature maps are doubled through bilinear interpolation and the maximum size feature maps use convolution whose stride is 2 to down-sample.Then the three feature maps have the same size and can be fused.In addition,this paper uses shortcut connection to connect different size of feature maps to obtain lost information from the larger feature maps.The model is evaluated on the autopilot international benchmark dataset KITTI and achieves 6%improvement compare to the SqueezeDet.The speed of the model reach 30fps on a GPU.