Design And Implementation Of Face Tracking Based On Multi Task Cascade Convolutional Neural Networks On ZYNQ

The performance of face tracking design based on embedded system is determined by two factors:(1)the detection efficiency and performance of the face detection algorithm;(2)the flexibility of the tracking platform based on embedded system.This design is dedicated to the innovation of face tracking design at these two aspects:About the face detection algorithm,the design utilizes the face detection algorithm based on multi-task cascade convolutional neural network which uses a complete cascading structure to predict the face and facial feature position from shallow to deep by a well-designed deep convolutional neural network.The performance of the algorithm is further improved through the internal correlation analysis between face detection and face alignment,it can also solve the low detection rate of the algorithm in complex facial expressions,facial occlusion and dim light effectively.In addition,the algorithm uses a new strategy named online difficult case mining in the process of model training,classifying the samples according to face approximation,and further improving the detection rate.In the process of transplanting algorithm to embedded system,the face model is transformed into NCNN model which is highly optimized on embedded system.About tracking platform based on embedded system,the hardware of design uses ZYNQ chip,which integrates ARM Cortex-A hard core and Xilinx 7 series FPGA programmable logic,it combines ARM's powerful transaction management function with FPGA parallel high-speed operation to enhanced the flexibility of the tracking platform greatly.At the same time,the design uses a programmable on-chip logic system(SOPC)as the system architecture which increases the reconfigurability of the system,not only saves time for design development and maintenance,but also provides convenience for future design extensions.This design has mainly completed the following tasks:(1)This article reviews a large number of data about intelligent monitoring systems,face detection algorithms and SOPC to obtain an optimized face tracking design scheme based on multi-task cascade convolutional neural network.Combined with the SOPC architecture to optimize the face tracking design.(2)The article introduces the principle of face detection algorithm used in the design,model transformation and deep learning framework in detail.The algorithm of the design is compared with other classic face detection algorithms to verify the superiority of this algorithm in detection performance.(3)This design completes the construction of hardware and software platform of face tracking design:the hardware mainly completes these contents:the selection of core board and related devices,the welding of and bottom board,the construction of steering engine platform;the software mainly builds the framework of embedded system and deep learning,and completes the training and transformation of model and the transplantation of algorithm into embedded system.(4)The design completes the realization of the face tracking function,and the experimental results show that the face detection algorithm proposed in this article combined with SOPC architecture has a higher detection rate and stronger performance,while still has a good detection effect in the complex facial expressions,facial occlusion and dim light environment.Through a series of specific data and analysis of the data,it is proved that the proposed optimization scheme of face tracking design in this article is executable.