Aiming at the situation that factors such as occlusion and angle change in image sequences are easy to cause head to miss the detection

a fast human head detection method is proposed which combined the cascade convolution object detection model MTCNN and kernel correlation filtering（ KCF） tracking model

namely MT-KCF head detection model. Firstly

the batch normalization is used to improve the MTCNN network

and the human head is detected by cascading. Secondly

the detected human head coordinate information is input into the KCF tracking model

and the head object is quickly and steadily tracked. Finally

to ensure that the human head is detected continuously and steadily

the human head is detected again using the detection model after tracking several frames. Experimental results show that the MT-KCF model has higher detection precision and faster detection speed in image sequences

and the average accuracy is 94. 85 %. The average speed in the 640 × 480 image sequences is 108 frames per second respectively.