|
|
|
|
|
<!DOCTYPE html><html><head>
|
|
|
|
|
|
<title>ReadMe</title>
|
|
|
|
|
|
<meta charset="utf-8">
|
|
|
|
|
|
<meta name="viewport" content="width=device-width, initial-scale=1.0">
|
|
|
|
|
|
|
|
|
|
|
|
<link rel="stylesheet" href="file:////home/wang/.vscode/extensions/shd101wyy.markdown-preview-enhanced-0.8.15/crossnote/dependencies/katex/katex.min.css">
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
<style>
|
|
|
|
|
|
code[class*=language-],pre[class*=language-]{color:#333;background:0 0;font-family:Consolas,"Liberation Mono",Menlo,Courier,monospace;text-align:left;white-space:pre;word-spacing:normal;word-break:normal;word-wrap:normal;line-height:1.4;-moz-tab-size:8;-o-tab-size:8;tab-size:8;-webkit-hyphens:none;-moz-hyphens:none;-ms-hyphens:none;hyphens:none}pre[class*=language-]{padding:.8em;overflow:auto;border-radius:3px;background:#f5f5f5}:not(pre)>code[class*=language-]{padding:.1em;border-radius:.3em;white-space:normal;background:#f5f5f5}.token.blockquote,.token.comment{color:#969896}.token.cdata{color:#183691}.token.doctype,.token.macro.property,.token.punctuation,.token.variable{color:#333}.token.builtin,.token.important,.token.keyword,.token.operator,.token.rule{color:#a71d5d}.token.attr-value,.token.regex,.token.string,.token.url{color:#183691}.token.atrule,.token.boolean,.token.code,.token.command,.token.constant,.token.entity,.token.number,.token.property,.token.symbol{color:#0086b3}.token.prolog,.token.selector,.token.tag{color:#63a35c}.token.attr-name,.token.class,.token.class-name,.token.function,.token.id,.token.namespace,.token.pseudo-class,.token.pseudo-element,.token.url-reference .token.variable{color:#795da3}.token.entity{cursor:help}.token.title,.token.title .token.punctuation{font-weight:700;color:#1d3e81}.token.list{color:#ed6a43}.token.inserted{background-color:#eaffea;color:#55a532}.token.deleted{background-color:#ffecec;color:#bd2c00}.token.bold{font-weight:700}.token.italic{font-style:italic}.language-json .token.property{color:#183691}.language-markup .token.tag .token.punctuation{color:#333}.language-css .token.function,code.language-css{color:#0086b3}.language-yaml .token.atrule{color:#63a35c}code.language-yaml{color:#183691}.language-ruby .token.function{color:#333}.language-markdown .token.url{color:#795da3}.language-makefile .token.symbol{color:#795da3}.language-makefile .token.variable{color:#183691}.language-makefile .token.builtin{color:#0086b3}.language-bash .token.keyword{color:#0086b3}pre[data-line]{position:relative;padding:1em 0 1em 3em}pre[data-line] .line-highlight-wrapper{position:absolute;top:0;left:0;background-color:transparent;display:block;width:100%}pre[data-line] .line-highlight{position:absolute;left:0;right:0;padding:inherit 0;margin-top:1em;background:hsla(24,20%,50%,.08);background:linear-gradient(to right,hsla(24,20%,50%,.1) 70%,hsla(24,20%,50%,0));pointer-events:none;line-height:inherit;white-space:pre}pre[data-line] .line-highlight:before,pre[data-line] .line-highlight[data-end]:after{content:attr(data-start);position:absolute;top:.4em;left:.6em;min-width:1em;padding:0 .5em;background-color:hsla(24,20%,50%,.4);color:#f4f1ef;font:bold 65%/1.5 sans-serif;text-align:center;vertical-align:.3em;border-radius:999px;text-shadow:none;box-shadow:0 1px #fff}pre[data-line] .line-highlight[data-end]:after{content:attr(data-end);top:auto;bottom:.4em}html body{font-family:'Helvetica Neue',Helvetica,'Segoe UI',Arial,freesans,sans-serif;font-size:16px;line-height:1.6;color:#333;background-color:#fff;overflow:initial;box-sizing:border-box;word-wrap:break-word}html body>:first-child{margin-top:0}html body h1,html body h2,html body h3,html body h4,html body h5,html body h6{line-height:1.2;margin-top:1em;margin-bottom:16px;color:#000}html body h1{font-size:2.25em;font-weight:300;padding-bottom:.3em}html body h2{font-size:1.75em;font-weight:400;padding-bottom:.3em}html body h3{font-size:1.5em;font-weight:500}html body h4{font-size:1.25em;font-weight:600}html body h5{font-size:1.1em;font-weight:600}html body h6{font-size:1em;font-weight:600}html body h1,html body h2,html body h3,html body h4,html body h5{font-weight:600}html body h5{font-size:1em}html body h6{color:#5c5c5c}html body strong{color:#000}html body del{color:#5c5c5c}html body a:not([href]){color:inherit;text-decoration:none}html body a{color:#08c;text-decoration:none}html body a:hover{color:#00a3f5;text-decoration:none}html body img{max-width:100%}html body>p{margin-top:0;margin-bottom:16px;word-wrap:break-word}html body>ol,html
|
|
|
|
|
|
/* Please visit the URL below for more information: */
|
|
|
|
|
|
/* https://shd101wyy.github.io/markdown-preview-enhanced/#/customize-css */
|
|
|
|
|
|
|
|
|
|
|
|
</style>
|
|
|
|
|
|
<!-- The content below will be included at the end of the <head> element. --><script type="text/javascript">
|
|
|
|
|
|
document.addEventListener("DOMContentLoaded", function () {
|
|
|
|
|
|
// your code here
|
|
|
|
|
|
});
|
|
|
|
|
|
</script></head><body for="html-export">
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
<div class="crossnote markdown-preview ">
|
|
|
|
|
|
|
|
|
|
|
|
<h1 id="sophon-bm1684x-算法耗时测试报告">sophon bm1684x 算法耗时测试报告 </h1>
|
|
|
|
|
|
<h2 id="1测试对象">1.测试对象 </h2>
|
|
|
|
|
|
<p>公司算能计算板卡，采用一个过遮挡的对地车辆视频进行跟踪+识别测试。<br>
|
|
|
|
|
|
<img src="img/image.png" alt="alt text"></p>
|
|
|
|
|
|
<p><strong>目标识别程序：</strong><br>
|
|
|
|
|
|
采用算能FP16精度模型，模型分辨率1280*1024</p>
|
|
|
|
|
|
<p><strong>目标传统跟踪算法：</strong><br>
|
|
|
|
|
|
采用跟踪基线（初始化为对地场景），KCF开3尺度，TLD抗遮挡模块开（OPENMP未开启）。</p>
|
|
|
|
|
|
<h2 id="2测试方法与结果">2.测试方法与结果 </h2>
|
|
|
|
|
|
<p>在算能板运行测试程序，并将识别、跟踪结果绘制在双路视频上，为了测试方便，双路均使用了相同的注入视频（视频规格1080p）。</p>
|
|
|
|
|
|
<p>为贴近实际产品，已有图像采集线程运行中。</p>
|
|
|
|
|
|
<h3 id="21-用例1">2.1 用例1 </h3>
|
|
|
|
|
|
<blockquote>
|
|
|
|
|
|
<p>单路传统跟踪</p>
|
|
|
|
|
|
</blockquote>
|
|
|
|
|
|
<p>仅运行单路传统算法跟踪程序时，单帧处理算法耗时如下，目标后半段过遮挡后有较大波动，其余大约在15ms左右。<br>
|
|
|
|
|
|
<img src="img/image-3.png" alt="alt text"></p>
|
|
|
|
|
|
<h3 id="22-用例2">2.2 用例2 </h3>
|
|
|
|
|
|
<blockquote>
|
|
|
|
|
|
<p>单路目标识别+单路目标传统跟踪</p>
|
|
|
|
|
|
</blockquote>
|
|
|
|
|
|
<p>目标识别接口自身可异步调用，因此直接同时运行即可。</p>
|
|
|
|
|
|
<p><img src="img/image-4.png" alt="alt text"><br>
|
|
|
|
|
|
同时运行目标识别+目标跟踪算法后，平均耗时提升到18ms，已较难满足50Hz，需要进一步优化程序。<br>
|
|
|
|
|
|
过遮挡时，整体波动更为明显。</p>
|
|
|
|
|
|
<h3 id="23-用例3">2.3 用例3 </h3>
|
|
|
|
|
|
<blockquote>
|
|
|
|
|
|
<p>按照双路目标识别+双路目标传统跟踪</p>
|
|
|
|
|
|
</blockquote>
|
|
|
|
|
|
<p>双路目标跟踪采用线程并行</p>
|
|
|
|
|
|
<p><img src="img/image-1.png" alt="alt text"></p>
|
|
|
|
|
|
<p>双路同时识别+跟踪与单路差异不大。双路识别跟踪效果如下：</p>
|
|
|
|
|
|
<p><img src="img/image-2.png" alt="alt text"></p>
|
|
|
|
|
|
<p>此时的CPU占用如下：<br>
|
|
|
|
|
|
<img src="img/image-5.png" alt="alt text"></p>
|
|
|
|
|
|
<h2 id="3结果说明">3.结果说明 </h2>
|
|
|
|
|
|
<p>算能板在2.3G主频下算力较为可观，能够基本满足50Hz对地需要，且双路跟踪与单路的耗时差异不大。<br>
|
|
|
|
|
|
整体上还需要进行一些针对性的优化将稳态耗时降低到15ms附近才能保证跟踪稳定。</p>
|
|
|
|
|
|
<p>对空算法整体资源消耗明显低于对地跟踪算法，且搜索与跟踪容易并行，可以充分利用多核优势，这里不再进行测试。</p>
|
|
|
|
|
|
<h2 id="4其他问题">4.其他问题 </h2>
|
|
|
|
|
|
<p>目前的算能板功耗和散热需要优化，双路跟踪仅需运行几分钟后，主频即减半，耗时提升一倍。</p>
|
|
|
|
|
|
|
|
|
|
|
|
</div>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
</body></html>
|
