To　improve　the　performance　of　traffic　sign　detection　and　recognition　systems　in　real　implementation　for　the　outdoor　challenging　environment,　we　propose　a　robust　traffic　sign　detection　algorithm　based　on　co-training　learning　methods　with　a　small　number　of　manually　labeled　initial　samples　(opposite　to　collect　all　possible　views)　in　this　paper.　With　consideration　on　the　various　appearances　of　different　traffic　signs　in　real　environment,　two　kinds　of　redundant　textual　descriptors　are　extracted　for　reinforcing　the　discrimination　ability　of　traffic　sign　detection　classifier　from　background.　First,　a　novel　traffic　sign　candidate　regions　extraction　method　is　used　based　on　probability　map　image　built　from　multiple　color-histogram　back-projection.　Secondly,　a　small　number　of　labeled　samples　are　used　to　train　two　classifiers　respectively:　one　is　AdaBoost　with　MB-LBP　(multi-block　local　binary　pattern)　features　and　the　other　is　SVM　(support　vector　machines)　with　HOG　(histograms　of　oriented　gradients)　features.　Then,　on　the　basis　of　co-training　semi-supervised　learning　framework,　the　newly　labeled　samples　with　higher　confidence　from　one　classifier　are　used　to　update　the　training　samples　of　the　other　one.　Because　of　the　constant　increment　of　each　training　samples,　the　performance　of　traffic　sign　detection　is　highly　improved　which　is　evaluated　intensively　in　the　results　of　our　experiment.