A small target detection algorithm working at a compressed domain was proposed for parallel compressive imaging systems to reduce the computational complexity by eliminating the process of image reconstruction. A mathematical model of the parallel compressive imaging system was used to capture measuring values of background and current frames. Then, the background measurements were updated according to a compressive sensing-mixture of Gaussians model (CS-MoG) to obtain the measurement values of the foreground. The cosine similarities between the measurements of current frame and the compressed target-location templates were calculated. And the local threshold and target area in the compressed domain were adopted to screen candidate targets. Finally, the effects of down-sampling rate, number of measurements, projection error and noise on the detection results were studied by simulation experiments. Experimental results show that large down-sampling rate and noise would decrease the detection performance, but the number of measurements to detection results has limited contribution. When 2 or 3 measurements are set, the operation time could be controlled while ensuring the detection performance. It suggests that the noise in the system should be controlled strictly because the noise effects the detection ability greatly. Furthermore, the proposed algorithm can achieve real-time target detection without any image reconstruction.