Study On Key Issues Of Traffic Sign Detection Under Natural Environments

Traffic sign recognition(TSR)system is an important subsystem of automatic driving system.Traffic sign detection(TSD)is an important foundation and key technology TSR.At present,TSD methods are mainly divided into three categories which are based on the color model,shape feature and machine learning.Detection methods based on the color model thresholding are a kind of simple and effective methods.However,the problem of signs missing detection due to signs interconnect exists in the methods.TSD based on fast radial symmetry transform(FRST)is one of the mainstream methods based on the shape feature,which has been widely concerned.However,the existing methods have some problems,such as difficulty in parameters setting for FRST,computational cost and complexity of algorithm,etc.Deep learning based sign detection has become a research hotspot.Nevertheless,the high risk of over-fitting arises when the number of the annotated images is relatively small in the training dataset.In addition,motion blur exists in some detected signs,which will seriouslyaffect the performance of TSR.To solve the above problems,researches are carried out on the key issues related to TSD systematically and deeply,such as separating interconnected signs,detecting traffic sign and deblurring traffic sign under natural environments.The main research contents and innovations are summarized as follows.(1)To solve the problems of performance degradation due to signs interconnected,blind separation of signs and high computational cost in the watershed transform based sign separation algorithm(WTSS),a watershed transform based signs adaptive separation algorithm(WTSAS)is proposed.In the proposed algorithm,interconnected candidate can be ed by the construction of a regional feature vector and the proposed judgment criterion of interconnected signs.In combination with the edge connection treatment using the morphological expansion and watershed transform segmentation,only the low-dimensional interconnected candidate is separated with ridge construction,effectively avoiding the blind separation.Experimental results show the proposed WTASS algorithm can discriminate and separate the interconnected signs adaptively,which will improve the separation performance and execution efficiency.To solve the possible problem of excessive separation of WTSAS,a novel separation method based on global and local curvature properties CSS(GLCP-CSS)corner detection method is proposed.In the proposed method,outline corners of interconnected candidate are effectively abstracted by introduction of the corner detection method based on GLCP-CSS.In combination with the proposed judgment criteria of the concave corner and matching criteria of separation corner pairs,the separation corner pairs are extracted and the signs are separated effectively.Experimental results show that the novel method overcomes the excessive separation of signs in WTSAS.The overall performance of the algorithm in the implementation efficiency and separation effect is superior to that of existing algorithms.(2)To solve the high computational cost,difficulty in parameters setting manually,and algorithm complexity in the existing methods based on FRST,a FRST based parameter-adaptive method is proposed.Through the proposed adaptive setting method of radius parameters in FRST,the problem of parameters setting is solved effectively.The FRST of the low-dimensional sign candidate is performed.In combination with the construction of center neighborhood of sign candidate and the proposed threshold criteria,the sign is detected effectively.Experimental results show that the novel algorithm can adaptively set the parameters in FRST and reduce the complexity and computational cost effectively through only processing the low dimensional candidates.To solve the limitation of the detection algorithms based on the shape feature,color model and traditional machine learning,a sign detection method based on the pseudo sample regularized Faster R-CNN deep model is proposed.In the proposed method,the Faster R-CNN model based target detection is introduced.To solve the high risk of over-fitting due to less training samples,a pseudo samples regularization method using data augment is proposed by adopting traffic signs and unannotated background sample provided in the training dataset.Experimental results show that the proposed method realizes the detection of various signs effectively,reduces the risk of over-fitting,and improves the detection performance.(3)To solve the motion blur of traffic signs and difficulty in achieving the reliable salient edge,a novel traffic sign deblurring algorithm based on exemplars is proposed.An exemplar dataset construction strategy and an exemplar matching strategy based on gradient information and entropy correlation coefficients(GECC)is constructed to lower computational cost and improve the matching accuracy.In the motion blur kernel estimation,the sparsity of the estimated kernel is further improved by introducing the L0.5-norm as the regularization.The theoretical analysis and experimental results verify the effectiveness of the proposed method in computational cost and accuracy of exemplar matching,and the sparsity of the estimated kernel.The overall performance of the method is superior to that of the existing methods.