Human head statistics is widely used in the construction of smart cities and has great market value. In order to solve the problem of missing pedestrian features and poor statistics results in a large-angle overlooking scene, in this paper we propose a human head statistics system that consists of head detection, head tracking and head counting, where the proposed You-Only-Look-Once-Head (YOLOv5-H) network, improved from YOLOv5, is taken as the head detection benchmark, the DeepSORT algorithm with the Fusion-Hash algorithm for feature extraction (DeepSORT-FH) is proposed to track heads, and heads are counted by the proposed cross-boundary counting algorithm based on scene segmentation. Specifically, Complete-Intersection-over-Union (CIoU) is taken as the loss function of YOLOv5-H to make the predicted boxes more in line with the real boxes. The results demonstrate that the recall rate and mAP@.5 of the proposed YOLOv5-H can reach up to 94.3% and 93.1%, respectively, on the SCUT_HEAD dataset. The statistics system has an extremely low error rate of 3.5% on the TownCentreXVID dataset while maintaining a frame rate of 18FPS, which can meet the needs of human head statistics in monitoring scenarios and has a good application prospect.

The convolutional neural network (CNN) has made great success in many fields, and is gradually being applied in edge-computing systems. Taking the limited budget of the resources in the systems into consideration, the implementation of CNNs on embedded devices is preferred. However, accompanying the increasingly complex CNNs is the huge cost of memory, which constrains its implementation on embedded devices. In this paper, we propose an efficient, pipelined convolution module based on a Brain Floating-Point (BF16) to solve this problem, which is composed of a quantization unit, a serial-to-matrix conversion unit, and a convolution operation unit. The mean error of the convolution module based on BF16 is only 0.1538%, which hardly affects the CNN inference. Additionally, when synthesized at 400 MHz, the area of the BF16 convolution module is 21.23% and 18.54% smaller than that of the INT16 and FP16 convolution modules, respectively. Furthermore, our module using the TSMC 90 nm library can run at 1 GHz by optimizing the critical path. Finally, our module was implemented on the Xilinx PYNQ-Z2 board to evaluate the performance. The experimental results show that at the frequency of 100 MHz, our module is, separately, 783.94 times and 579.35 times faster than the Cortex-M4 with FPU and Hummingbird E203, while maintaining an extremely low error rate.