Research On Traffic Sign Detection Based On Convolutional Neural Network

In recent years,self-driving cars have drawn increasing interest from both academia and industry.As an important part of the self-driving cars’ environmental percep-tion system,traffic sign detection system is also getting more and more attention.It is extremely important to correctly detect and classify the traffic signs.These years convo-lutional neural networks(CNN)have achieved a number of successful applications in the field of computer vision.Existing CNN based detection algorithms such as Faster R-CNN,YOLO and SSD,have achieved quite good performance on some general purpose object detection dataset.However,their performance is still far from the practical application of traffic sign detection in the wild.This paper focuses on applying and improving CNN based traffic sign detector.The contributions are as follows:(1)We proposed a novel feature transferring algorithm that transferring a,digit clas-sifier’ s features to a traffic sign detector.Due to the extreme imbalance of training samples,the drastic variant of the scale and the complexity of the background,the fea-tures learned by the traffic sign detector are often sub-optimal.Especially when the traffic sign contains digit.To address this problem,we firstly train a digit classifier network,then using the digit classifier’s high level features as an additional supervise signal to the traffic sign detector.After the feature transferring,our traffic sign detector gets more discriminative features and significantly outperforms the baseline algorithm.(2)We improved the CNN based traffic sign detector by applying a feature weight-ing mechanism.Current CNN based traffic sign detectors usually do not consider the effectiveness of features.However,for a particular traffic sign category only part of the features have positive effects,others are useless or even have negative effects.Based on the existing image classification model,we improved the traffic sign detector by applying a feature weighting mechanism.When using Faster R-CNN as the basic detector,we introduce a feature weighting branch after the RoI Pooling layer.This feature weighting branch can produce a weighting vector,which models the importance of each channel.Feature weighting algorithm can be interpreted as a soft feature selection mechanism,which encourages useful features and suppresses useless features.Extensive experiments validated the effectiveness of our improving.