基于尺度感知CNN的实时车辆检测算法

Real-time Vehicle Detection Algorithm Based on Scale-aware Convolutional Neural Network

交通场景视频中的车辆尺度范围变化较大,不同空间尺度的车辆的实例表现出不同特征,导致特征的类内方差较大,影响检测方法的识别率。针对现有基于CNN的算法中卷积特征对尺寸不具有鲁棒性的问题,提出一种基于尺度感知的卷积神经网络(SAVD)车辆检测算法。该算法采用尺度感知的ROI池化层来维护小尺寸对象的原始结构。针对较大尺寸变化的类内距离超过单个网络的表示能力的问题,内置分支决策子网络来最小化特征的类内距离。实验证明,该算法在准确率上显著提高,对不同尺寸实例具有鲁棒性,上述轻量级技术提高检测速度,具有较好的实时性。

The video in the traffic scene contains vehicles with a large variance of scales,different scales of vehicles exhibit different features,which results in large intra-category variance in features,may affect the performance of the detection methods.Aiming at the problem that the convolution features of existing CNN-based algorithms are not robust to scale,a scale-aware convolutional neural network(SAVD)vehicle detection algorithm is proposed.The algorithm proposes a scale-aware ROI pooling layer to maintain the original structure of small-sized objects.The problem that the intra-class distance for larger size changes exceeds the representation ability of a single network,which introduces multiple built-in decision subnetworks to minimize intraclass distance of features.Experiments show that the algorithm is significantly improved in accuracy and robust to different sizes instance the above lightweight technology improves the detection speed in real time.