基于DSP的全景视频多目标实时检测

摘要

Because panoramic video has a 360 degree viewing angle, it is quite suitable for the multi-object detection in intelligent video surveillance systems. First, this paper introduces the DSP-based system hardware architecture. Second, the panoramic image unrolling algorithm and the MHOEI multi-object detection algorithm are given. Then the look-up method of block table on DSP platform for panoramic image unrolling and the resolution reduction method for MHOEI are designed respectively. Finally, the optimized scheme is implemented on DSP processor. Experimental results show that it achieves 120 f/s processing speed for unrolling the panoramic video resolution from 720×576 into 896×180 on the selected DM6437 processor. And after combined with the multi-object detection, the average speed is up to 35 f/s, which meets real-time requirements. Meanwhile, it can detect moving and motionless objects in panoramic video effectively.%全景视频由于具有360°的视角，非常适合用于智能视频监控中的多目标检测。首先介绍了以 DSP 处理器为核心的系统处理平台；接着介绍了全景图像的展开算法和MHOEI多目标检测算法；然后针对全景展开设计了适合DSP平台的分块查表展开方法，针对MHOEI设计了降分辨率处理方法；最后在DSP处理器上进行了优化实现。实验表明，在所选的DM6437处理器上，将720×576大小的全景图像展开为896×180大小的目标图像，展开帧率可达120帧每秒；再经过多目标检测后能达到平均35帧每秒，满足实时要求，同时能有效检测出全景视频中的运动和静止目标。