Design And Implementation Of Object Detection Algorithms In Embedded Systems

In recent years,deep learning object detection algorithms have been widely used in the fields of automatic driving,industrial quality detection,intelligent safety and etc.However,there are many problems in the related algorithms,such as the number of neural network,the number of network parameters and the amount of computation,which makes it a great challenge to complete the real-time object detection in the embedded system with limited hardware resources and can not meet the needs of time-sensitive services.Therefore,this thesis focuses on the embedded system,and optimizes the design and implementation of the algorithm based on MobileNetV2-SSD real-time object detection.Firstly,in order to speed up and lighten the object detection algorithm from the network structure,this thesis analyzes the network structure of the MobileNetV2-SSD object detection algorithm and puts forward a structural optimization scheme that completely replaces the standard convolution in MobileNetV2-SSD object detection algorithm with deep separable convolution,and makes a quantitative analysis of the calculation and parameters of the scheme.Secondly,in order to speed up the calculation of the model in the inference process,this thesis adjusts the structure of the trained model,designs the fusion of convolution layer and BN(Batch Normalization)layer,and integrates the convolution calculation and BN layer calculation which need two steps into matrix multiplication which only needs one step.Then,this thesis starts from the network parameters to compress the model size and speed up the model inference.Firstly,the model parameters were optimized,and the unsigned integer 8-bit quantization scheme was adopted in the model inference stage to alleviate the problem of insufficient computation and storage resources of the embedded system.Quantization design of matrix multiplication was carried out for the convolution layer and the activation layer.Secondly,in order to further accelerate the speed of object detection,this thesis puts forward an optimization scheme combining Kalman filter prediction with object detection algorithm,that is,using Kalman filter algorithm to directly predict object detection frame,thus reducing the frequency of model inference of the object detection algorithm,and obtaining lower average delay and power consumption of object detection.At the same time,the simulation results show that the algorithm can complete the object detection with good detection accuracy.Finally,in order to verify the optimization algorithm of embedded object detection designed in this thesis,a real-time embedded object detection system is designed and implemented.The hardware of the system is based on the iMX6 processor and Google tensor computing unit,while software relies on GStreamer streaming media framework and embedded model inference API TensorFlow Lite.In this system,the performance of the optimization algorithm is tested.Test results show that the algorithm optimization design proposed in this thesis can realize real-time object detection with low power consumption,low delay,and high performance in the embedded system.