Study On A Traffic Sign Detection Algorithm Based On Full Convolutional Network

Thanks to the development of convolutional neural networks,object detection has progressed rapidly in recent years.Its applications have been extended to various industry fields.Current popular object detection algorthims are based on deep convolutional neural network.An object detection algorithm based on one-step method and full convolutional network is proposed.The algorithm maps the targets to the feature maps of multiple output scales through the relation between step size and receptive field.Then Targets could be predicted from feature map directly.Two strategies are used to improve the prediction effect of the one-step method in this process.The first is to form redundant predictions for the same object through different quantization methods.The second is to use the area threshold on each scale to form an implicit constraint,so that each scale feature map will be forced to predict target within its area range.Because it is unrealistic to label the target coordinates to an integer,and the coordinates need to be divided by the step size to map to the feature map,the coordinates of the target on the feature map after mapping are floating point values.However,the coordinates of the floating-point values cannot correspond to specific pixels,so that the pixels on the feature map cannot be divided into positive / negative examples.Therefore,the floating point value needs to be quantized as an integer to correspond to a specific pixel on the feature map.Thess points are positive examples and responsible for predicting current Ground Truth.By combining different quantization strategies during quantization,a ground truth is formed into multiple positive examples to predict it,which improves the effect of object detection.When using multi-scale prediction strategy,according to the characteristics of each scale feature,a larger-sized feature map should be responsible for predicting a target with a smaller area,and a smallersized feature map should predict a target with a larger area.Constraints on predicting targets are formed by specifying different area ranges of targets for distinct feature maps of different scales during training.Based on a high-performance multi-card GPU platform,this algorithm is implemented and applied to traffic sign data sets.The correspondence between the target and the feature map pixels is formed by point mapping.This mapping make each feature map’s pixel a positive example,so that the entire network can be trained in an end-to-end manner which improves the training efficiency.In summary,this methed is a traffic sign detection algorithm with better comprehensive performance.Experiments has revealed that the above two strategies effectively improve the detection effect.Meanwhile accuracy and speed proved to be realiable when applying this algorithm to the detection of traffic signs.