Abstract: Considering the effects of stalling, quality switching, content characteristics and other factors, which will be directly reflected in the distorted video, a client-oriented mobile video perceptual quality assessment model was proposed. The mapping model between the distorted video and the mean opinion score (MOS) was established based on the idea of "deep feature extraction+regression" instead of characterizing and measuring each influencing factor. First, ResNet-TSM network was constructed to extract the deep spatial-temporal features of each distorted video segmentation. Second, LargeVis algorithm was used to reduce the dimensionality of the extracted deep features, and simultaneously improving the representation and discriminative capabilities of the features. Afterward, the score of each video segment was obtained by modeling the long-term dependence of the video by using the bidirectional gated recurrent unit. The temporal mean pooling was adopted to aggregate the scores of each segment to obtain the overall video score. The experimental results on the WaterlooSQoE-Ⅲ and LIVE-NFLX-Ⅱ datasets show that the proposed model can achieve a higher prediction accuracy.