An image sequence compressed sensing algorithm by minimizing prediction errors was proposed for high speed camera image compression in real-time. First, an original image was compressed only by a projection matrix on the encoder side. The observed vector obtained by compressing was transferred to the decoder through a channel. Then, motion estimation and motion compensation were performed on correlated images on the decoder side, and a prediction image was generated in this way. Furthermore, the prediction error image which is the difference between original image and prediction image was reconstructed by compressed sensing. Finally, the reconstruction of prediction error image was improved by an iterative procedure, until the difference between two consecutive reconstruction results was smaller than a predetermined threshold. Therefore, the original image was reconstructed by the prediction error image. Experiments by CR-GEN0-H6400 camera from DALSA indicate that the proposd algorithm can compress 1 000 frame/s images in real-time, and image reconstruction result is improved by 2-6 dB at least as compared with that of independent reconstruction. The proposed algorithm can compress high speed camera images in real-time, and can reconstruct the images in high quality.