—Very few benchmark exists for assessing pattern detection and recognition in streams in general and for gesture processing in particular. We propose a dedicated benchmark based on the construction of isolated gestures and gesture sequences datasets. This benchmark is associated to a general assessment methodology for streaming processing which first consists in labelling the stream according to some heuristics (that can be optimized on training data) and then aligning the ground truth labelling with the predicted one. 6 pattern recognition models (including DTW, KDTW, HMM, HCRF and SVM) have been accordingly evaluated using this benchmark. It turns out that the regularized kernelized version of DTW measure (KDTW) associated to a SVM is quite efficient, comparatively to the other models, for detecting and recognizing continuous gestures in streams.