Traffic Sign Detection Based on Histograms of Oriented Gradients and Boolean Convolutional Neural Networks
Zhitao Xiao, Zhenjie Yang, Lei Geng, Fang Zhang · 2017
State-of-the-art methods for traffic signs detection based feature extraction have got a high recall rate, but the detection rates are not ideal for some mistakenly detected. In this work, this paper presents a method of traffic signs detection based on HOG and Boolean Convolutional Neural Networks (HOG-BCNN). A cascade classifier is trained based on HOG to detect the candidate regions of traffic signs. These regions as proposal windows input a special CNN, which is like a Boolean logic. BCNN is utilized to eliminate the false detected regions in the proposal traffic sign windows, which is connected to the cascade classifier as the last stage. In BCNN, the stochastic gradient descent method is exploited to minimize the error in the discriminant of true and false, which likes a supervisor guides the network how to reduce failures. The proposed method has been evaluated on the real environment, and experiments demonstrate that it is sufficient to attain high detection accuracy and ensure the efficiency.