Vision-based　traffic　accident　detection　is　one　of　the　challenging　tasks　in　intelligent　transportation　systems　due　to　the　multi-modalities　of　traffic　accidents.　The　first　challenging　issue　is　about　how　to　learn　robust　and　discriminative　spatio-temporal　feature　representations.　Since　few　training　samples　of　traffic　accidents　can　be　collected,　sparse　coding　techniques　can　be　used　for　small　data　case.　However,　most　sparse　coding　algorithms　which　use　norm　regularisation　may　not　achieve　enough　sparsity.　The　second　challenging　issue　is　about　the　sample　imbalance　between　traffic　accidents　and　normal　traffic　such　that　detector　would　like　to　favour　normal　traffic.　This　study　proposes　a　traffic　accident　detection　method,　including　a　self-tuning　iterative　hard　thresholding　(ST-IHT)　algorithm　for　learning　sparse　spatio-temporal　features　and　a　weighted　extreme　learning　machine　(W-ELM)　for　detection.　The　ST-IHT　algorithm　can　improve　the　sparsity　of　encoded　features　by　solving　an　norm　regularisation.　The　W-ELM　can　put　more　focus　on　traffic　accident　samples.　Meanwhile,　a　two-point　search　strategy　is　proposed　to　adaptively　find　a　candidate　value　of　Lipschitz　coefficients　to　improve　the　tuning　precision.　Experimental　results　in　our　collected　dataset　have　shown　that　this　proposed　traffic　accident　detection　algorithm　outperforms　other　state-of-the-art　methods　in　terms　of　the　feature＇s　sparsity　and　detection　performance.