For the real-time problem of deep learning algorithm in multi-target tracking, a multi-target tracking algorithm based on MobileNet is proposed in this paper. Using the principle that deepwise separable convolution of MobileNet can compress deep network models, the YOLOv3 backbone network is replaced by MobileNet, but multiscale predictions are remained. This reduces effectively the amount of parameters by factorizing a standard convolution into a depthwise convolution and a pointwise convolution. For the detected bounding box information, the tracking task is performed by the Deep-SORT algorithm. The experimental results show that the processing speed is improved by nearly 50% under the condition of maintaining the tracking result.