Pedestrian detection has become the core technology that security, intelligent video surveillance, and traffic statistics of people in the scenic area depend on. The latest object detection methods such as FastRegions with Convolution Neural Network (FastRCNN), Faster RCNN, Single Shot Multibox Detector (SSD), Deformable Part Models (DPM) are currently the classic algorithms for object detection. However, these algorithms pay more attention to detect the whole pedestrians. In large scenes, pedestrians have different postures and some of them are occluded frequently. Only modeling the position of the pedestrian’s body and grasping the local features of the pedestrians can achieve accurate positioning. The FasterRCNN deep network prototype is adopted, a detection model is built for pedestrian heads, head features in different directions are extracted at the same time, and a spatial pyramid pooling layer is added to ensure the detection rate. These can effectively solve the partial occlusion problem of pedestrians in large scenes and clearly show the general flow direction of pedestrians. The proposal is more conducive to the flow statistics than the ordinary head estimation.