Automatic crowd counting has attracted widespread concern in the field of video surveillance. In recent years, the Convolutional Neural Network(CNN) has achieved miraculous results in crowd counting. However, current research based on deep learning mainly concentrates on high-performance PC to count the people with a single still picture. The network model has huge computational resources consuming due to its large amount of parameters and complex network structure, which is difficult to deploy in actual surveillance video crowd counting system. Therefore, the deep learning method is adopted to realize the real-time crowd counting on the embedded platform by pruning and compressing the network model and using TensorRT to accelerate the model inference. The proposed crowd counting algorithm achieves a balance between accuracy and speed with Mean Absoulte Error(MAE) of 21.6 and average Frames Per Second(FPS) of 22. Its performance on the embedded platform can approach the real-time result.