脈沖多普勒雷達

by Braden Riggs and George Williams (gwilliams@gsitechnology.com)

Braden Riggs和George Williams(gwilliams@gsitechnology.com)

In the world of data science the industry, academic, and government sectors often collide when enthusiasts and experts alike, work together to tackle the challenges we face day-to-day. A prime example of this collaboration is the Israeli Ministry of Defense Directorate of Defense Research & Development (DDR&D)’s MAFAT challenges. A series of data science related challenges with real-world application and lucrative prize pools. In the program’s own words:

在數據科學世界中，當發燒友和專家都共同努力應對我們日常面臨的挑戰時，行業，學術界和政府部門經常發生沖突。 以色列國防部國防研究與發展局(DDR＆D)的MAFAT挑戰就是這種合作的主要例證。 一系列與數據科學相關的挑戰，包括現實應用和豐厚的獎池。 用程序本身的話來說：

The goal of the challenge is to explore the potential of advanced data science methods to improve and enhance the IMOD current data products. The winning method may eventually be applied to real data and the winners may be invited to further collaborate with the IMOD on future projects.- MAFAT Competition Coordinators

挑戰的目標是探索先進數據科學方法的潛力，以改善和增強IMOD當前數據產品。 獲獎方法可能最終會應用于真實數據，獲獎者可能會被邀請在未來的項目上與IMOD進一步合作。- MAFAT競賽協調員

Given the recent inception of the program, there haven’t been many challenges yet, however, there are expected to be a variety of challenges ranging from complicated Natural Language Processing puzzles to computer-vision related endeavors.

鑒于該程序是最近啟動的，因此還沒有很多挑戰，但是，預計會出現各種各樣的挑戰，從復雜的自然語言處理難題到計算機視覺相關的工作。

One such challenge, their second one made available thus far, caught my eye. It involves creating a model for classifying living, non-rigid objects that have been detected by doppler-pulse radar systems. The challenge, “MAFAT Radar Challenge — Can you distinguish between humans and animals in radar tracks?” implores competitors to develop a model that can accurately distinguish humans from animals based on a spectrum of radio signals recorded from various doppler-pulse radar sites on various days. If you are interested in participating I recommend visiting the challenge site before reading on.

這樣的挑戰之一，到目前為止已經提供的第二個挑戰引起了我的注意。 它涉及創建一個模型，以對多普勒脈沖雷達系統已檢測到的活的非剛性物體進行分類。 挑戰“ MAFAT雷達挑戰-您能區分雷達軌道中的人與動物嗎？ 懇請競爭對手開發一種模型，該模型可以根據在不同日子從各個多普勒脈沖雷達站點記錄的無線電信號頻譜，準確地將人與動物區分開。 如果您有興趣參加，我建議您訪問 在繼續閱讀之前先挑戰網站 。

那么，我們正在處理什么樣的數據？我們需要了解什么？ (So what kind of data are we working with and what do we need to know about it?)

An example of the data included for the competition split by Animal/Human and High/Low Signal-Noise-Ratio. The I/Q matrices have been converted into spectrograms for visualization, and the doppler readings have been added in white. As you can see there are some differences present in the files. Images provided by MAFAT. Reposted with Author’s permission. 比賽數據包括動物/人類和高/低信噪比。 I / Q矩陣已轉換為頻譜圖以進行可視化，并且多普勒讀數已添加為白色。 如您所見，文件中存在一些差異。 圖片由 MAFAT 提供 。 經作者許可重新發布。

The key to developing an accurate and competitive model is to first understand the data, how it was sourced, and what it is missing. Included with the competition is 5 CSV files containing the metadata, and 5 pickle files (serializing Python object structure format) containing doppler readings that track the object’s center of mass and slow/fast time readings in the form of a standardized I/Q matrix.

開發準確而具有競爭力的模型的關鍵是首先了解數據，數據來源和缺失內容。 競賽中包括5個包含元數據的CSV文件 ，以及5個 包含多普勒讀數的 pickle文件 (序列化Python對象結構格式) ，它們以標準化I / Q矩陣的形式跟蹤對象的質心和慢/快時間讀數 。

Before we go any further it is worth breaking down a few key concepts relating to signals and the specific types of data collected. The signal readings that make up the dataset fall into two levels of quality, High Signal to Noise Ratio, and Low Signal to Noise Ratio. This reading, High SNR and Low SNR divides the set into two levels of quality, one with high clarity that hasn’t been heavily tainted by a noise generating process, and one with low clarity that has had aspects such as weather impact the quality of the reading.

在進一步研究之前，有必要分解一些與信號和所收集數據的特定類型有關的關鍵概念。 構成數據集的信號讀數分為兩個質量級別，即高信噪比和低信噪比 。 該讀數分為高信噪比和低信噪比兩類，將質量分為兩個級別，一個級別的高清晰度沒有受到噪聲生成過程的嚴重影響，而另一個級別的低清晰度卻受到天氣等因素的影響。閱讀。

KF6HI. Reposted with Author’s permission.KF6HI 。 經作者許可重新發布。

You might be wondering why we would even choose to include low SNR readings given the impact noise has on the data, however to my surprise this data is actually quite valuable when developing an effective model. Real-life is messy, and the true reading one might expect to see will not always be high quality, hence it is important to make sure our model is adaptive and geared towards a range of data readings, not just the highest quality ones. Furthermore, we are working with a limited amount of data (which we will explore in-depth below) and hence want to utilize everything at our disposal for training the model.

您可能想知道為什么考慮到噪聲對數據的影響，我們為什么甚至選擇包括低SNR讀數，但是令我驚訝的是，在開發有效模型時，該數據實際上非常有價值。 現實生活是一團糟，人們可能期望看到的真實讀數并不總是高質量的，因此，重要的是要確保我們的模型具有自適應性，并且適合各種數據讀數，而不僅僅是高質量的讀數。 此外，我們正在處理數量有限的數據(我們將在下面深入探討)，因此希望利用我們掌握的所有信息來訓練模型。

Another series of concepts worth understanding is the notion of an I/Q matrix and what a doppler reading entails. An I/Q matrix consists of an N x M matrix, in our case a 32 x 128 matrix, that stores the slow and fast signal readings as cartesian elements, where “I” represents the real part and “M” represents the imaginary part. You can picture each row of this matrix as representing a signal pulse from the source, and each column of this matrix representing a reading for returning radio waves that have bounced off objects or targets in the direction of interest. The time between pulses is “slow time” and the time between readings of said pulses is considered “fast time”, if you are still confused or further interested I highly recommend you follow this link for more information.

值得理解的另一系列概念是I / Q矩陣的概念以及多普勒讀數的含義。 I / Q矩陣由N x M矩陣(在我們的示例中為32 x 128矩陣)組成，該矩陣將笛卡爾元素的慢和快信號讀數存儲為笛卡爾元素，其中“ I”代表實部，“ M”代表虛部。 您可以將矩陣的每一行表示為代表來自源的信號脈沖，并將矩陣的每一列表示為用于返回沿感興趣方向從物體或目標反彈的無線電波的讀數。 脈沖之間的時間為“慢時間”，而所述脈沖之間的時間間隔為“快速時間”，如果您仍然感到困惑或對此有進一步的興趣，我強烈建議您點擊此鏈接以獲取更多信息。

A visualization of fast time relative to slow time. In our case, the I/Q matrix would have 32 rows and 128 columns. Image by Author.快速時間相對于慢時間的可視化。 在我們的情況下，I / Q矩陣將具有32行和128列。 圖片由作者提供。

Also included in the dataset, separate from the I/Q matrix, is the doppler burst readings. Consisting of one row of 128 readings the doppler burst can be used to track an object’s speed and direction of travel. Much like how the sirens on a police car change sound as the car drive past you, the doppler effect relates to the range in wavelength characteristics of objects in motion. By bouncing radio signals off objects of interest we can see how the radio waves change shape and hence infer a number of parameters about the object of interest such as speed, direction, and acceleration.

與I / Q矩陣分開的數據集中還包括多普勒猝發讀數。 由128個讀數的一行組成，多普勒脈沖串可用于跟蹤物體的速度和行進方向。 就像警車上的警笛聲一樣，當汽車駛過您時，多普勒效應與運動物體的波長特性范圍有關。 通過將無線電信號彈離目標物體，我們可以看到無線電波如何改變形狀，從而推斷出有關目標物體的許多參數，例如速度，方向和加速度。

Great, now that we have a bit of terminology under our belt it is time to discuss the five file pairs provided for the competition. These file pairs, whilst in the same format, differ from each other greatly and form five distinct sets:

太好了，現在我們有了一些專業術語，現在該討論為比賽提供的五對文件了。 這些文件對雖然格式相同，但彼此之間有很大差異，并形成五個不同的集合：

• Training set: As the name describes, the training set consists of a combination of human and animal, with high and low SNR readings created from authentic doppler-pulse radar recordings.

  訓練集：顧名思義，訓練集由人和動物組成，具有由真實的多普勒脈沖雷達記錄創建的高和低SNR讀數。

  Training set: As the name describes, the training set consists of a combination of human and animal, with high and low SNR readings created from authentic doppler-pulse radar recordings.6656 Entries

  訓練集：顧名思義，訓練集由人和動物組成，具有從真實的多普勒脈沖雷達記錄創建的高和低SNR讀數。 6656個條目

• Test set: For the purposes of the competition, a test set is included to evaluate the quality of the model and rank competitors. The set is unlabeled but does include a balanced mix of high and low SNR.

  測試集：出于競爭目的，其中包括一個測試集，用于評估模型的質量并為競爭對手排名。 該集合未標記，但包含高和低SNR的平衡混合。

  Test set: For the purposes of the competition, a test set is included to evaluate the quality of the model and rank competitors. The set is unlabeled but does include a balanced mix of high and low SNR.106 Entries

  測試集：出于競爭目的，其中包括一個測試集，用于評估模型的質量并為競爭對手排名。 該集合未標記，但包含高和低SNR的平衡混合。 106條目

• Synthetic Low SNR set: Using readings from the training set a low SNR dataset has been artificially created by sampling the high SNR examples and artificially populating the samples with noise. This set can be used to better train the model on low SNR examples.

  合成低信噪比集：使用訓練集中的讀數，通過對高信噪比示例進行采樣并用噪聲人工填充樣本來人工創建低信噪比數據集。 此集合可用于在低SNR實例上更好地訓練模型。

  Synthetic Low SNR set: Using readings from the training set a low SNR dataset has been artificially created by sampling the high SNR examples and artificially populating the samples with noise. This set can be used to better train the model on low SNR examples.50883 Entries

  合成低信噪比集：使用訓練集中的讀數，通過對高信噪比示例進行采樣并用噪聲人工填充樣本來人工創建低信噪比數據集。 此集合可用于在低SNR實例上更好地訓練模型。 50883條目

• The Background set: The background dataset includes readings gathered from the doppler-pulse radars without specific targets. This set could be used to help the model better distinguish noise in the labeled datasets and help the model distinguish relevant information from messy data.

  背景集：背景數據集包括從多普勒脈沖雷達收集的，沒有特定目標的讀數。 該集合可用于幫助模型更好地區分標記數據集中的噪聲，并幫助模型將相關信息與混亂數據區分開。

  The Background set: The background dataset includes readings gathered from the doppler-pulse radars without specific targets. This set could be used to help the model better distinguish noise in the labeled datasets and help the model distinguish relevant information from messy data.31128 Entries

  背景集：背景數據集包括從多普勒脈沖雷達收集的，沒有特定目標的讀數。 該集合可用于幫助模型更好地區分標記數據集中的噪聲，并幫助模型從混亂數據中區分相關信息。 31128個條目

• The Experiment set: The final set and possibly the most interesting, the experiment set includes humans recorded by the doppler-pulse radar in a controlled environment. Whilst not natural this could be valuable for balancing the animal-heavy training set provided.

  實驗裝置：最終裝置，也許是最有趣的裝置，該實驗裝置包括多普勒脈沖雷達在受控環境中記錄的人類。 這雖然不自然，但對于平衡提供的大量動物訓練集可能很有??價值。

  The Experiment set: The final set and possibly the most interesting, the experiment set includes humans recorded by the doppler-pulse radar in a controlled environment. Whilst not natural this could be valuable for balancing the animal-heavy training set provided.49071 Entries

  實驗裝置：最終裝置，也許是最有趣的裝置，該實驗裝置包括多普勒脈沖雷達在受控環境中記錄的人類。 這雖然不自然，但對于平衡提供的大量動物訓練集可能很有??價值。 49071條目

As I have already alluded to, the training set isn’t populated with a satisfactory amount of data points. This constitutes the challenge, generating a sufficient amount of data to train the model on, from the supplementary synthetic, background, and experimental sets. This challenge is further exacerbated by the imbalance of the data.

正如我已經提到的，訓練集中沒有填充令人滿意的數據點。 這就構成了挑戰，需要從補充的合成，背景和實驗集中生成足夠數量的數據來訓練模型。 數據不平衡進一步加劇了這一挑戰。

With such a small dataset it is important to ensure the data is balanced and unbiased as this can lead to significant misinterpretations of the set by the model, and small inconsistencies can get extrapolated into significant errors.

使用如此小的數據集，重要的是要確保數據平衡且無偏見，因為這可能導致模型對集合的嚴重誤解，并且小的不一致性可能會推斷出重大錯誤。

Image by Author圖片作者

The first key imbalance is the difference between the number of high and low SNR tracks. As you can see from the adjacent graph there are almost two thousand more low SNR data points than high SNR.

第一個關鍵失衡是高和低SNR磁道數之間的差異。 從相鄰的圖表中可以看到，低SNR數據點比高SNR多了近兩千。

|

|

Image by Author圖片作者

The second key imbalance is between the number of Humans and Animals in the dataset. Clearly, with such a significant difference the model might become biased towards predicting animal instead of human, since this prediction would net a high accuracy for little effort on the model’s part.

第二個關鍵的不平衡是數據集中的人類和動物數量之間。 顯然，由于存在這種顯著差異，因此該模型可能會偏向于預測動物而不是人類，因為這種預測將為模型方面付出很少的努力而獲得很高的準確度。

|

|

Image by Author圖片作者

Both of these disparities cause significant issues when building the model. If we take a closer look at the relationship between signal quality and target type we see that the majority of animals have low SNR readings and the majority of humans have high SNR readings. Whilst this may seem minor, extrapolated over a number of training intervals our model may make the mistake of conflating a cleaner signal with that of a human, and a noisy signal with that of an animal.

建立模型時，這兩個差異都會導致嚴重問題。 如果我們仔細研究信號質量和目標類型之間的關系，我們會發現大多數動物的SNR讀數較低，而大多數人的SNR讀數較高。 盡管這似乎很小，但在許多訓練間隔中推斷出來，我們的模型可能會犯錯誤，將較干凈的信號與人的信號混淆，而將噪聲信號與動物的信號混淆。

|

|

基準模型和初始印象： (The Baseline Model and Initial Impressions:)

Interestingly enough, along with the data provided, a baseline model was included for the competitors. This model serves as an example of how the final submission should be formatted as well as providing a relative starting point for competitors. So what is the baseline model?

有趣的是，連同所提供的數據，還包括了針對競爭對手的基線模型。 該模型是如何格式化最終提交文件以及為競爭對手提供相對起點的示例。 那么基線模型是什么？

The MAFAT challenge organizers decided to start strong by beginning with Convolutional Neural Network (CNN), a form of artificial intelligence designed for computer vision problems. The model takes an input image and weights parameters based on their importance in discerning the final result, which in our case would be an animal or a human. This particular CNN has two convolutional layers, followed by two max-pooling layers, which again is followed by two “dense” layers, before finally being activated by a ReLU function and regularized with a Sigmoid function. This is better visualized with a diagram:

MAFAT挑戰賽組織者決定從強大的卷積神經網絡(CNN)開始 ，這是為計算機視覺問題設計的一種人工智能形式。 模型根據輸入在識別最終結果中的重要性來獲取輸入圖像和權重參數，在我們的例子中，該結果將是動物或人類。 這個特定的CNN具有兩個卷積層 ，然后是兩個最大池化層 ，然后又是兩個“密集”層 ，最后由ReLU函數激活并由 Sigmoid函數進行正則化。 使用圖表可以更好地將其可視化：

MAFAT. Reposted with Author’s permission.MAFAT 。 經作者許可重新發布。

As you can see in the above diagram we start with the 126x32 I/Q matrix. This matrix, along with 15 others, are aligned, and the first convolution of training happens, of which the result is altered and resized to a different dimensionality. Eventually, the model concludes with a single value, a number somewhere between 0 and 1 where the closer to 0 the more likely the signal is an animal, and the closer to 1 the more likely the signal is human. It is alright if you don’t understand the logic or the terminology behind this baseline model, these techniques are quite elaborate and if I were to go into detail this blog would be twice as long. If you are interested this link goes into more detail.

如上圖所示，我們從126x32 I / Q矩陣開始。 該矩陣與其他15個矩陣對齊，并且發生了第一次訓練卷積，其結果被更改并調整為不同的維度。 最終，模型以單個值結束，該數字介于0到1之間，其中數字越接近0，表示該信號越可能是動物，而數字越接近1，則表示該信號更可能是人類。 如果您不了解此基準模型背后的邏輯或術語，那也沒關系，這些技術都非常詳盡，如果我要詳細介紹，那么此博客的時間將是原來的兩倍。 如果您有興趣，請訪問此鏈接 。

In addition to the model, the baseline attempt includes a few other noteworthy strategies for increasing the accuracy of prediction. As discussed earlier the training set is heavily imbalanced, to help amend this discrepancy the training set is supplemented with more data from the experiment set. This is to help the CNN understand and recognize human patterns within the data and will ideally lead to a higher level of accuracy. In our own attempt, we trained the model without changing the baseline structure, and validated (scored the accuracy of the model) on a sample of the training data withheld from the model. The results are visualized below:

除模型外，基線嘗試還包括其他一些值得注意的策略，可以提高預測的準確性。 如前所述，訓練集嚴重失衡，為了幫助糾正這種差異，訓練集還添加了來自實驗集的更多數據。 這是為了幫助CNN了解和識別數據中的人為模式，并且理想情況下將導致更高的準確性。 在我們自己的嘗試中，我們在不更改基線結構的情況下對模型進行了訓練，然后對從模型中保留的訓練數據樣本進行了驗證(對模型的準確性進行評分)。 結果顯示如下：

Results of baseline model graphed. Image by Author.繪制基線模型的結果。 圖片由作者提供。

As you can see from the results the model performed perfectly on the training data, and almost perfectly on the validation set. For a baseline model, this is pretty impressive, right? Well as it turns out, by the admission of MAFAT themselves, the baseline model doesn’t perform well on the test set, averaging only a 75% accuracy. Given the scope of the project and the technology they are trying to produce, 75% simply won’t cut it. Hence we have to go back to the drawing board to figure out how we can create a more accurate model.

從結果中可以看出，該模型在訓練數據上表現完美，在驗證集上表現幾乎完美。 對于基準模型，這是非常令人印象深刻的，對嗎？ 事實證明，通過MAFAT本身的接受，基線模型在測試集上的表現不佳，平均準確性僅為75％。 考慮到項目范圍和他們試圖生產的技術，75％根本不會削減。 因此，我們必須回到制圖板上，找出如何創建更準確的模型。

什么不起作用，我們可以看到一種模式嗎？ (What isn’t working and can we see a pattern?)

So now that we understand how the baseline model works we need to understand what kind of mistakes the model is making on the test data. The best way to understand these patterns and the mistakes made by the model is to visualize the data, although this is easier said than done. Because of the high dimensionality of the data, it can be hard to visualize and understand in a meaningful way. Luckily for us, there is a solution to this problem, T-distributed Stochastic Neighbor Embedding for high dimensional data, also known as TSNEs. A TSNE is essentially its own machine learning algorithm for non-linear dimension reduction. It works by constructing a probability distribution over the different pairings of data where higher probabilities can be imagined as pairings of higher similarity. As the TSNE function continues it repeats this process, slowly predicting dimensionality until it reaches a stage where it is digestible to the human brain. Our code for producing the TSNE, along with the baseline notebook can be found here. In our case, we extracted the vector representation of the spectrogram using the final layer of the network before classification and computed the TSNE on the resulting vector.

因此，既然我們了解了基準模型的工作原理，我們就需要了解該模型在測試數據上犯了什么樣的錯誤。 理解這些模式和模型所犯錯誤的最好方法是可視化數據，盡管說起來容易做起來難。 由于數據的高維度，可能很難以有意義的方式可視化和理解。 幸運的是，對于這個問題，有一個解決方案，即針對高維數據的T分布隨機鄰居嵌入，也稱為TSNE。 TSNE本質上是其自己的用于非線性降維的機器學習算法。 它通過在不同數據對上構建概率分布來工作，其中較高的概率可以想象為較高相似性的對。 隨著TSNE功能的繼續，它會重復此過程，慢慢預測維數，直到達到人腦可消化的階段。 我們用于生產TSNE的代碼以及基準筆記本 可以在這里找到。 在我們的案例中，我們提取了 在分類之前使用網絡的最后一層對頻譜圖進行矢量表示，并在生成的矢量上計算TSNE。

Because of the stochastic nature of the algorithm, TSNE’s look different every time they are computed, however, they are useful for pointing out noteworthy clusters of similar data. Computing the TSNE for our model produces the following plot where:

由于該算法具有隨機性，因此每次計算時，TSNE的外觀都會有所不同，但是，它們對于指出相似數據的值得注意的簇很有用。 為我們的模型計算TSNE會產生以下圖，其中：

Green = animalBlue = humanRed = incorrect prediction in the validation setTeal = location of a test set value

綠色=動物藍色=人類紅色=驗證集中的預測不正確Teal =測試集值的位置

TSNE graph. Image by Author.TSNE圖。 圖片由作者提供。

As you can see there are some pretty significant clusters of animals and a few clusters of humans. Because there are fewer humans in the training set the human clusters are less apparent when compared to the animal clusters. As indicated by the red points there are a few areas where the model makes incorrect predictions. This is noteworthy because it appears as though the red points form two distinct clusters themselves, suggesting that the majority of incorrectly predicted points are close to two separate epicenters. What is also noteworthy is that there are a significant number of teal points that also fall in these regions, which explains why the baseline model is only scoring around ~75%, because the model would be incorrectly predicting these points.

如您所見，這里有一些非常重要的動物群和一些人類群。 由于訓練集中的人較少，因此與動物群相比，人群的明顯性較低。 如紅色點所示，模型在一些區域進行了錯誤的預測。 這是值得注意的，因為它看起來好像紅點本身形成了兩個不同的簇，這表明大多數錯誤預測的點都靠近兩個單獨的震中。 還值得注意的是，在這些區域中也有大量的藍綠色點，這解釋了為什么基線模型僅得分在?75％左右，因為模型會錯誤地預測這些點。

It also appears that the test set is relatively spread out not forming as clear of a center and being relatively even between animals and humans, although we can’t know this for sure as we don’t possess the labels for points at those locations.

似乎測試集也相對分散，沒有形成清晰的中心，并且在動物和人之間相對均勻，盡管我們不能確切知道這一點，因為我們在那些位置沒有點的標簽。

下一步： (Where to next:)

It can be hard to know which direction to take the project. Photo by Javier Allegue Barros on Unsplash很難知道該項目的發展方向。 Javier Allegue Barros在Unsplash上拍攝的照片

Given this information, there are a number of different strategies we can explore for boosting the overall quality of the model or in creating a different model altogether. In an ideal world, we would have a larger training dataset, this would be a great solution to the problem as the more points we have to train on, the more chance the model has of understanding a pattern that can lead to the correct classification of the red clusters above. Unfortunately, this isn’t an option and we are limited to the data provided or any data we can gather from external sources. This seems like a good place to start because the distribution is so unbalanced between humans, animal, low SNR, and high SNR. By developing a better distribution of data, be that from the auxiliary sets provided, or from some external source, we can retrain the model and see how the results improve. Depending on the performance of the baseline model on a more balanced dataset, we can then move forward towards creating an improved model.

有了這些信息，我們可以探索許多不同的策略來提高模型的整體質量或完全創建不同的模型。 在理想的世界中，我們將擁有一個更大的訓練數據集，這將是一個很好的解決方案，因為我們必須訓練的點越多，該模型就越有機會理解可以正確分類的模式。上面的紅色簇。 不幸的是，這不是一種選擇，我們僅限于提供的數據或我們可以從外部來源收集的任何數據。 這似乎是一個不錯的起點，因為人，動物，低SNR和高SNR之間的分布非常不平衡。 通過開發更好的數據分布，無論是從提供的輔助集中還是從某些外部來源，我們都可以重新訓練模型并查看結果如何改善。 根據基線模型在更平衡的數據集上的性能，然后我們可以朝著創建改進的模型前進。

As I write this now some competitors have already scored accuracies greater than 95%. A leader board of competitors and their scores can be found here. This is a multipart series with more updates to come as we proceed through the competition.

在撰寫本文時，一些競爭對手的準確度已超過95％。 一個 競爭對手排行榜及其分數可在此處找到。 這是一個分為多個部分的系列，隨著比賽的進行，將會有更多更新。

資料來源和其他閱讀： (Sources and Additional Reading:)

IQ Modulation. (n.d.). Retrieved August 13, 2020, from https://www.keysight.com/upload/cmc_upload/All/IQ_Modulation.htm?cmpid=zzfindnw_iqmod

智商調制。 (nd)。 于2020年8月13日從https://www.keysight.com/upload/cmc_upload/All/IQ_Modulation.htm?cmpid=zzfindnw_iqmod檢索

Saha, S. (2018, December 17). A Comprehensive Guide to Convolutional Neural Networks — the ELI5 way. Retrieved August 13, 2020, from https://towardsdatascience.com/a-comprehensive-guide-to-convolutional-neural-networks-the-eli5-way-3bd2b1164a53

Saha，S.(2018年12月17日)。 卷積神經網絡綜合指南-ELI5方法。 于2020年8月13日從https://towardsdatascience.com/a-comprehensive-guide-to-convolutional-neural-networks-the-eli5-way-3bd2b1164a53檢索

Understanding I/Q Signals and Quadrature Modulation: Radio Frequency Demodulation: Electronics Textbook. (n.d.). Retrieved August 13, 2020, from https://www.allaboutcircuits.com/textbook/radio-frequency-analysis-design/radio-frequency-demodulation/understanding-i-q-signals-and-quadrature-modulation/

了解I / Q信號和正交調制：射頻解調：電子教科書。 (nd)。 于2020年8月13日從https://www.allaboutcircuits.com/textbook/radio-frequency-analysis-design/radio-frequency-demodulation/understanding-iq-signals-and-quadrature-modulation/檢索

What is I/Q Data? (n.d.). Retrieved August 13, 2020, from http://www.ni.com/tutorial/4805/en/

什么是I / Q數據？ (nd)。 于2020年8月13日從http://www.ni.com/tutorial/4805/en/檢索

All images used are either created by myself or used with the explicit permission of the authors. Links to the author’s material are included under each image.

所有使用的圖像要么由我自己創建，要么在作者的明確許可下使用。 每個圖像下方都包含指向作者資料的鏈接。

翻譯自: https://towardsdatascience.com/training-a-model-to-use-doppler-pulse-radar-for-target-classification-2944a312148c

脈沖多普勒雷達

總結

以上是生活随笔為你收集整理的脉冲多普勒雷达_是人类还是动物？ 多普勒脉冲雷达和神经网络的目标分类的全部內容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網站內容還不錯，歡迎將生活随笔推薦給好友。