Fast detection of multiple objects in traffic scenes with a common detection framework

IEEE Transactions on Intelligent Transportation Systems. 2015

本文使用一个通用框架来检测三个东西:车、交通信号标志、骑自行车的人。
aggregated channel features + Shrinkage version of AdaBoost

系统大的框架如下:

A. Object Subcategorization
对于同一类的物体,如果彼此相差很大,我们将其分成若干个子类,对每个子类训练一个检测器进行检测。
1) Visual Features 我们使用 ACF特征,比HOG效果要好
2) Geometrical Features 几何特征是很重要的特征,这里我们使用 3D orientation,Aspect-ratio,Truncation level,Occlusion index

3) Clustering 这里我们使用 normalized spectral clustering 进行分子类

B. Feature extraction
1) Aggregated channel features (ACF):LUV color channels ,
Gradient magnitude channel,Gradient histogram channels
2) Spatially pooled features: Covariance matrix–Spatially pooled covariance
Local Binary Pattern–Spatially pooled LBP

C. Supervised learning
Fast detection of multiple objects in traffic scenes with a common detection framework_第1张图片
使用 Bootstrapping 训练较难的样本

D. Post-processing
1) Calibration of confidence scores
2) Non-maximum suppression (NMS)
3) Fusion of detection results

IV. E XPERIMENTS

Fast detection of multiple objects in traffic scenes with a common detection framework_第2张图片
Fast detection of multiple objects in traffic scenes with a common detection framework_第3张图片
Fast detection of multiple objects in traffic scenes with a common detection framework_第4张图片

你可能感兴趣的:(Fast detection of multiple objects in traffic scenes with a common detection framework)