Traffic Sign Detection And Recognition Algorithm Research Based On Monocular Vision

With the increase of car ownership, traffic safety has become an unignoredproblem and the research of Intelligent Transportation System has attracted theattention of the countries. As an important corpus of Intelligent Tranportation System,the auxiliary driving system must realize the lane line detection, vehicle andpedestrian detection, obstacle detection, traffic sign recognition, etc. Traffic signautomatic detection and recognition system could effectively reduce traffic violationcaused by negligence and greatly decrease the traffic accident rate to ensure efficientand safe driving. Traffic sign detection and recognition system based on monocularvision is fulfilled through the following steps: the road scene images are captured byon-board monocular camera firstly, image procession algorithms are used toaccomplish traffic sign preliminary detection secondly, pattern recognition techniqueis applied to traffic sign recognition lastly. In this paper, coarse segmentationalgorithm based on color feature, shape detection algorithm, feature extractionclassification method and template matching algorithm were analyzed and studied.In terms of traffic sign detection, the various kinds of coarse segmentationalgorithms based on color feature were first introduced and compared. Fixed thresholdsegmentation result was easily influenced by external factors such as light change, soan adaptive threshold segmentation method based on HSI was proposed to overcomethe drawback. Secondly, labeling based on region growing algorithm was applied toexcluding some non-sign region and effectively segmenting multiple connected signregion. Then, several typical edge detection methods were analyzed and a simplemethod based on Log edge detector and two-way scanning filter was presented to getthe outside edge. Combined with color features, circle similarity parameter, rectanglesimilarity parameter and triangle similarity parameter were used to divide traffic signsinto four categories: red round, blue round, blue rectangle and yellow triangle.In terms of traffic sign recognition, traffic signs candidate blocks were firstpreprocessed: bicubic interpolation for candidate block size normalized processingand two-way scanning filter combined with color features for segmenting thecandidate block inside characteristics. Secondly, by using Hu invariant moments and8-adjacency connected region number of traffic signs candidate block insidecharacteristics image, preliminary classification was accomplished to decrease the number of templates and improve recognition efficiently. Finally, traffic signs wererecognized successfully using the improved weighted templates and discriminantfunctions based on distance and angle and the experiment results demonstrated a highaccuracy.