自动驾驶场景下实时目标检测模型评估计算平台的研究与实现论文设计

摘要随着人工智能迅猛发展和传感器成本不断降低，自动驾驶技术进入了快速发展阶段，其通过传感器获取环境信息，由计算系统感知环境并以此作出相应决策。由于环境感知具有高复杂性，实现面向该场景的计算存储平台仍有诸多挑战：1）感知计算所依赖的传感器数据海量多样，当数据接口升级时，通信困难；2）感知算法研究多关注提高模型训练效果，而忽略模型推断过程；3）感知中 3D 目标检测与 2D 目标检测相比，检测评估基准仍有较大差异。基于上述问题，本文提出了一种关于快速帮助算法研究人员迭代开发的方案，使平台在分布式处理并存储传感器数据时进行数据兼容扩展并以此实现一站式目标检测评估。主要研究工作：1）基于 TensorRT 的模型推断。为加快深度学习的模型推断，采用 TensorRT 加速器加快推断过程，提高算法研发效率；2）Benchmark 模型评估机制。基于 KITTI 的模型评估指标，结合 3D 目标检测数据的特点，设计 benchamrk 模型评估机制，并与模型推断相结合，实现一站式目标检测的评估。3）数据兼容扩展。采用 Protobuf 在分布式存储计算过程中进行兼容扩展，解决通信问题。关键词：分布式存储，Protobuf，TensorRT，KITTI，benchamrkiAbstractWith the rapid development of artificial intelligence and the continuous reduction of sensor costs, autonomous driving technology has entered a stage of rapid development. It obtains environmental information through sensors, and the computing system senses the environment and makes corresponding decisions. Due to the high complexity of environmental perception, there are still many challenges in implementing a computing storage platform oriented to this scenario: 1) The sensor data on which perceptual computing depends is massive and diverse, the data interface is upgraded, and communication is difficult; 2) the research of perception algorithms focuses more on improving model training Effect, while ignoring the model inference process; 3) Compared with 2D target detection, 3D target detection in perception still has a large difference in the evaluation benchmark.Ba...