Fast Stereo-RCNN三维目标检测算法

Fast Stereo-RCNN 3D Target Detection Algorithm

智能机器人、无人驾驶等技术的飞速发展在引领产业变革的同时对环境感知技术提出了新的挑战.基于视觉的三维目标检测算法的性能有了大幅提升,直逼激光雷达,但检测速度离产业实时性需求还有较大差距,成为产业发展的瓶颈之一.鉴于此,本文提出一种基于Stereo-RCNN的Fast Stereo-RCNN三维目标检测算法,用单支路网络检测三维框的多个角点重构三维中心点,轻量区域生成网络固化三维关键点,二分支关键点检测网络锐化目标辨识能力,双层特征融合网络缩短低层特征到高层特征的传递路径.实验结果表明,Fast Stereo-RCNN在检测精度提高的同时检测时间从0.3秒/帧降到了0.11秒/帧,提高了1.72倍.

The rapid development of intelligent robot,unmanned driving and other technologies has brought new challenges to environmental awareness technology while leading the industrial revolution.The performance of 3D target detection algorithm based on vision has been greatly improved,reaching the level of lidar.However,the detection speed still lags far behind the real-time requirements of the industry,which has become one of the bottlenecks of industrial development.In view of this,this paper proposes a Fast Stereo-RCNN based on Stereo-RCNN 3D target detection algorithm,with one branch network to detect the three dimensional frame multiple angular point to reconstruct 3D center,three dimensional network curing light regions,two branch point detection network sharpening target recognition,double feature fusion network shorten the low-level features to high-level transfer path.The experimental results show that the detection time of Fast Stereo-RCNN decreases from 0.3 seconds/frame to 0.11 seconds/frame while the detection accuracy is improved,an increase of 1.72 times.