Abstract: Due to the limitations of image acquisition equipment and poor acquisition conditions, it is a challenge to obtain high-quality images in reality, especially for video images. However, the existing instance segmentation algorithms can hardly handle low-resolution (LR) videos. Moreover, since existing complicated instance segmentation models can be barely applied to mobile devices in practical applications. Accordingly, the proposed method develops an efficient and lightweight instance segmentation model built upon MobileNet. At the same time, an improved super-resolution coding-based network (SCN) algorithm for low-resolution video is proposed as preprocessing. In addition, the motion vector is employed as post-proposing to predict the inter-frame mask. Experimental results have demonstrated that the proposed algorithm could be easily transplanted to embedded platforms in LR real-time street view dataset thanks to its remarkably low memory cost and high precision.