Face Detection(2) - AISTARVISION dragonboard 410c

最新推荐文章于 2018-12-07 18:05:46 发布

原创最新推荐文章于 2018-12-07 18:05:46 发布 · 916 阅读

0 ·

CC 4.0 BY-SA版权

文章标签：

#aistarvision #DragonBoard 410c #camera #machine vision #MIPI CSI

DragonBoard 410c 专栏收录该内容

12 篇文章

订阅专栏

本文深入探讨了人脸检测（Face Detection）技术的关键参数及其应用场景。详细介绍了Face Detection软件如何分析场景并检测其中的人脸，包括其在不同角度下的人脸识别能力、最小人脸尺寸设置以及在预览和录制过程中连续运行的配置。此外，还提到了为提高人脸识别数量所做的定制化架构改进。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

衔接Face Detection(1)，我们这篇博文share further parameters of FD core feature.

The FD feature software analyzes a scene and detects the faces in the scene. On detecting the faces, FD sends relevant information to the other features as follows,

1. Face 3A ( AF, AEC, AWB ).

2. Display face ROIs ( region of interest ) in preview.

3. Miscellaneous ( Face Noise Reduction and so on ).

FD can detect faces in the following angles:

1. 360° of roll angles.

2. -90° to 90° of yaw angles.

3. -30° to 30° of pitch angles.

FD is typically configured to run continuously during preview for both front and rear cameras. It can also be configured to run during video preview and recording. 在上篇博文中演示的demo，就是FD during preview for rear camera.

The FD core only accepts the following dimensions; the VFE must downscale to these dimensions:

1. WVGA – 854 x 480

2. VGA – 640 x 480

3. WQVGA– 427 x 240

4. QVGA – 320 x 240

Minimum face size settings

1. 0x0 – Set minimum face size to 20 pixels

2. 0x1 – Set minimum face size to 25 pixels

3. 0x2 – Set minimum face size to 32 pixels

4. 0x3 – Set minimum face size to 40 pixels

对于Face Detect，Qualcomm VFE会有单独一路output stream - Viewfinder stream给到FD core consumption.

目前我们在customize FD architecture，以提高单路stream的人脸识别数量，先定一个小目标，从5提高到10。同时也开始基于双路dual camera进行machine vision和deep learning视觉与图像方面的开发。