The invention discloses an improved multi-target detection method based on sparse representation. The method comprises steps of: extracting the HOG characteristic of a training sample and performing sparse representation on the HOG characteristic by using a K-SVD algorithm to obtain a over-complete dictionary D; traversing an image to be detected by using a sliding window and a multi-scale strategy, extracting the HOG characteristic of each image sub block, and performing sparse representation based on the over-complete dictionary D; preliminarily positioning multiple targets by using the difference of reconstruction errors in combination with an adaptive threshold setting method; and finally designing a window merging strategy to accurately detect the multiple targets. The window merging strategy provided in the invention differentiates multiple target positioning windows by using the central positions and the sizes of the windows in order that the windows accurately correspond to respective targets, and merges the multiple superposed positioning windows corresponding to the same target so as to guarantee the robustness and accuracy detection of the targets.