Concerning the problem of poor detection of human body keypoints based on video streams and possible motion blur after video stream slicing, an improved RetinaNet-CPN network is proposed to detect the keypoints, avoiding the interference of motion-blurred images after slicing and improving the detection accuracy of the keypoints. After the video stream is sliced, the improved RetinaNet network is first used to detect all the people in the picture and perform fuzzy detection on each target frame. The target frame larger than the threshold is deblurred, and finally, the keypoints are extracted with the CPN network with the attention mechanism. After the IOU function of RetinaNet to measure the difference between the predicted frame and the real frame is changed into DIOU, the target detection AP increases by nearly 3% in the simulation experiment. For blurry pictures, the blur kernel estimated with the spectrum feature of uniform linear motion is slightly different from the actual blur kernel, and the original clear picture can be restored after the deblurring. At the same time, the attention mechanism is adopted to assign reasonable weights to each channel and feature layer, which increases the CPN detection AP by nearly 1% and the AR by 0.5%.