Porting And Evaluation Of ADAS Object Detection Algorithm For RISC-Ⅴ Platform

Object detection is one of the important perception tasks in the Advanced Driving Assistant System(ADAS),which is a typical Artificial Intelligence Internet of Things(AIo T)application.Limited by hardware resources,AI-based object detection methods in ADAS are difficult to implement in vehicle-mounted systems.As a new instruction set architecture,Reduced Instruction Set Computing V(RISC-Ⅴ)is very suitable for the development of AIo T and is expected to promote the implementation of ADAS applications.However,there is currently a lack of practical experience and evaluation data for ADAS applications porting to the RISC-Ⅴ platform.The ADAS-oriented RISC-Ⅴ CPU design lacks relevant references.In order to explore the feasibility of porting ADAS-oriented target detection algorithms to the RISC-Ⅴ platform,a set of evaluation schemes are proposed.First,the performance changes of the four ADAS detection algorithms after transplanting to RISC-Ⅴ are evaluated,and the speed and accuracy of down-sampling pooling or full convolution in the detection network are compared.Second,the basic integer instruction set for RISC-Ⅴ has different implementations of RV32 E and RV32(64)I,while RV32 E only uses the first 16 of the 32general-purpose registers in RV32(64)I,which is more suitable for embedded environments.However,there is currently no experimental environment that supports RV32 E,so the evaluation of the frequency of register usage by the detection network is launched to provide a basis for the selection of RISC-Ⅴ’s integer instruction set.The above experiment is based on the RISC-Ⅴ instruction simulator Spike,and the data set uses COCO-ADAS,a subset of Common Objects in Context(COCO).Experimental results show that the accuracy of the model will not change after being transplanted to the RISC-Ⅴ platform.Compared with the pooling implementation of downsampling,the number of RISC-Ⅴ instructions for forward inference of the full convolutional network has dropped by 32.3%,and the accuracy has dropped by 13.1%.Therefore,using convolution to downsampling has a speed advantage and is more suitable for scenes that focus on realtime performance such as ADAS.The evaluation results of the register usage ratio show that the calculation of the object detection network uses 81.4% of the first 16 registers and only 18.6% of the last 16 registers,indicating that the first 16 registers are used more frequently.Therefore,the ADAS-oriented RISC-Ⅴ CPU can choose to use RV32 E,which uses the first 16 registers,as the basic integer instruction set to reduce the complexity of hardware implementation.