In this paper, we present a simple event-oriented algorithm for detection of
pixel-size signals with a known frequency, by the novel technology of an event
camera. In addition, we analyze the ability of the algorithm to filter out the
desired periodic signals from random fluctuations. We demonstrate this ability
and show how the algorithm can distinguish, during twilight, between the
signals of a streetlight that flicker with frequency of 100 Hz, and sun glitter
originating from windows in far-away buildings in the field of view.
An asynchronous event-based algorithm for periodic signals
周期的信号に対する非同期イベントベースアルゴリズム
0.73
David El-Chai Ben-Ezra, Ron Arad, Ayelet Padowicz, Israel Tugendhaft
David El-Chai Ben-Ezra, Ron Arad, Ayelet Padowicz, Israel Tugendhaft
0.45
Abstract In this paper, we present a simple event-oriented algorithm for detection of pixel-size signals with a known frequency, by the novel technology of an event camera.
In addition, we analyze the ability of the algorithm to filter out the desired periodic signals from random fluctuations.
さらに,ランダムな変動から所望の周期信号をフィルタリングするアルゴリズムの能力を解析する。
0.67
We demonstrate this ability and show how the algorithm can distinguish, during twilight, between the signals of a streetlight that flicker with frequency of 100 Hz, and sun glitter originating from windows in far-away buildings in the field of view.
In particular, using frame-based cameras, one needs to sample the signal at a rate higher than double the desired frequency (due to the Nyquist criterion).
This approach works nicely when the desired signal is not too fast, and not too short.
このアプローチは、所望の信号が速すぎず、短すぎない場合にうまく機能する。
0.74
However, if the frequency of the signal is higher than ~1 kHz, then sampling it with a frame-based camera and analyzing it with this approach can become quite cumbersome, and if the signal is too short, one might miss it altogether due to the inherent camera dead time.
We got so used to this approach that it takes a bit of thinking in order to realize that using the deep theory of Fourier transform for this mission sounds like using a 5 kilo hammer in order to knock a nail.
The asynchronous bio-inspired paradigm of event cameras (see [2] and [4] for review) offers an approach that sounds much more natural and intuitive: to check whether the time difference between adjacent events in a certain pixel corresponds to the desired frequency.
Adapting this approach to an event camera, one can easily surpass the 1 kHz limit, without missing a short signals, as event-cameras do not have a dead time.
In this paper, we use the notion of a "time-surface" (see [2] and references 21, 80, 109114 therein) and present an asynchronous event-based building block algorithm to distinguish between signals of a given frequency and random flickering, based on this intuitive approach.
At the end of the paper we demonstrate the algorithm performance, using the presented analysis, and show how it distinguishes between the periodic signals of street lights flickering at 100 Hz, and other random signals during twilight, when many objects in the field of view are flickering as a result of sun glittering.
Description of the algorithm For the sake of simplicity, we consider only the positive events.
アルゴリズムの説明 単純さのため、ポジティブな出来事のみを考慮する。
0.61
However, depending on the application, one can take the negative events instead, or consider both polarities.
しかし、アプリケーションによっては、代わりに負のイベントを取るか、両方の極性を考えることができる。
0.64
Therefore, the input of the algorithm is the list of positive events generated by the event camera
したがって、アルゴリズムの入力は、イベントカメラが生成した陽性事象のリストである。
0.81
𝐿 = {(𝑡𝑖, 𝑥𝑖, 𝑦𝑖) | 𝑖 = 1, … , 𝑘}
𝐿 = {(𝑡𝑖, 𝑥𝑖, 𝑦𝑖) | 𝑖 = 1, … , 𝑘}
0.42
where 𝑡𝑖 is the 𝑖-th event timestamp, (𝑥𝑖, 𝑦𝑖) is its pixel location and 𝑘 is the number of events.
ti は i 番目のイベントタイムスタンプで (xi, yi) はピクセル位置、k はイベントの数です。 訳抜け防止モード: ti が i 番目のイベントタイムスタンプ、 (xi, yi) がそのピクセルの位置です。 そしてkはイベントの数です。
0.79
We define the two following variables, which we keep constant along the algorithm:
アルゴリズムに沿って定数を保つ2つの変数を定義します。
0.68
1. 𝛿 ≔ the period of the signal we are looking for.
1. δ は私たちが探している信号の周期である。
0.84
2. 𝜖 ≔ the error we take into account in the period of the signal.
2. ε ... 信号の周期で考慮する誤差。
0.53
Typical value for this variable should be the expected rise time of the signal.
典型的な値 この変数は信号の上昇時間として期待される。
0.81
As mentioned in the introduction, we use the notion of "time surface".
紹介で述べたように、私たちは"時間表面"という概念を使います。
0.75
Namely, in the initialization of the algorithm, we define a 2D array which we initialize its entries arbitrarily to be a some negative number, smaller than – 𝛿 − 𝜖.
The output of the algorithm is a sub-list 𝐿𝑝𝑒𝑟𝑖𝑜𝑑𝑖𝑐 ⊆ 𝐿, sorted from 𝐿, using a for-loop going along 𝑖 = 1, … , 𝑘, implementing the following orders:
アルゴリズムの出力は、i = 1, ... , k に沿って進むforループを用いて、L からソートされたサブリスト L 周期 L である。 訳抜け防止モード: アルゴリズムの出力は、サブ-リスト L 周期 L, i = 1 に沿って進むfor-ループを使って、L からソートする。 …, 𝑘, 以下の命令を実行します
0.85
1. If |𝑡𝑖 − 𝑇𝑆𝑥𝑖,𝑦𝑖 − 𝛿| < 𝜖, then add (𝑡𝑖, 𝑥𝑖, 𝑦𝑖) to 𝐿𝑝𝑒𝑟𝑖𝑜𝑑𝑖𝑐.
This building block of the algorithm by itself does not decide whether we have found the desired periodic signal.
アルゴリズム自体のこの構築ブロックは、所望の周期信号が見つかったかどうかを判断しない。
0.81
This is done afterwards, by picking the pixels that the number of events in 𝐿𝑝𝑒𝑟𝑖𝑜𝑑𝑖𝑐 associated to them are surpassing a threshold, as described below.
Algorithm analysis We want to explain how to use the algorithm in order to distinguish between the desired periodic signal, and all the other random flickering.
In order to do that, after implementing the algorithm on the recording, for each pixel, we evaluate the number 𝑚 of events reported by the pixel during the recording, and the number 𝑛 of events related to the pixel in 𝐿𝑝𝑒𝑟𝑖𝑜𝑑𝑖𝑐 .
𝛿+𝜖 𝛿−𝜖 However, as we do not know the value of 𝜆, this probability should be weighted by the probability of having 𝑚 events given a Poisson distribution with parameter 𝜌 = 𝜆 ⋅ 𝑇.
𝛿+𝜖 𝛿−𝜖 しかし、λ の値が分かっていないので、この確率はパラメータ ρ = λ ⋅ t のポアソン分布が与えられる m 事象を持つ確率によって重み付けされるべきである。
0.48
Hence, given the fact that we only know the number 𝑚 and not the parameter 𝜆, the probability of an event to be added to 𝐿𝑝𝑒𝑟𝑖𝑜𝑑𝑖𝑐 is given by
したがって、数 m だけを知っていてパラメータ λ を知らないという事実を考えると、l periodic に追加される事象の確率は与えられる。
0.78
𝑃(𝛿, 𝜖, 𝑚, 𝑇) = ∫ 𝑃̃(𝛿, 𝜖, 𝜆) ⋅
𝑃(𝛿, 𝜖, 𝑚, 𝑇) = ∫ 𝑃̃(𝛿, 𝜖, 𝜆) ⋅
0.42
∞ 𝜌𝑚𝑒−𝜌 𝑚!
∞ 𝜌𝑚𝑒−𝜌 𝑚!
0.38
⋅ 𝑑𝜌 0 ∞ = 𝑇 ∫
⋅ 𝑑𝜌 0 ∞ = 𝑇 ∫
0.43
0 (𝜆𝑇)𝑚(𝑒−𝜆(𝑇+𝛿−𝜖) − 𝑒−𝜆(𝑇+𝛿+𝜖))
0 (𝜆𝑇)𝑚(𝑒−𝜆(𝑇+𝛿−𝜖) − 𝑒−𝜆(𝑇+𝛿+𝜖))
0.38
𝑚! ⋅ 𝑑𝜆. Then, using repeatedly the method of integration by parts, one gets that
𝑚! ⋅ 𝑑𝜆. そして、部品による積分の手法を繰り返すと、それを得られる。
0.52
𝑃(𝛿, 𝜖, 𝑚, 𝑇) = 𝑇𝑚+1 (
𝑃(𝛿, 𝜖, 𝑚, 𝑇) = 𝑇𝑚+1 (
0.49
1 (𝑇 + 𝛿 − 𝜖)𝑚+1 −
1 (𝑇 + 𝛿 − 𝜖)𝑚+1 −
0.46
1 (𝑇 + 𝛿 + 𝜖)𝑚+1).
1 (𝑇 + 𝛿 + 𝜖)𝑚+1).
0.46
Hence, omitting the first event reported by the pixel, the probability of at least 𝑁 events out of 𝑀 = 𝑚 − 1 to be added to 𝐿𝑝𝑒𝑟𝑖𝑜𝑑𝑖𝑐 is given by
したがって、ピクセルによって報告された最初の事象を省略すると、M = m − 1 から少なくとも N 個の事象が L 周期に付加される確率は与えられる。
0.69
𝑄(𝛿, 𝜖, 𝑚, 𝑇, 𝑁) = ∑ (
𝑄(𝛿, 𝜖, 𝑚, 𝑇, 𝑁) = ∑ (
0.42
𝑚−1 𝑙=𝑁 𝑚 − 1
𝑚−1 𝑙=𝑁 𝑚 − 1
0.34
𝑙 ) 𝑃(𝛿, 𝜖, 𝑚, 𝑇)𝑙(1 − 𝑃(𝛿, 𝜖, 𝑚, 𝑇))
𝑙 ) 𝑃(𝛿, 𝜖, 𝑚, 𝑇)𝑙(1 − 𝑃(𝛿, 𝜖, 𝑚, 𝑇))
0.42
𝑚−1−𝑙 . Remark: In Appendix 2, we show that in certain cases, a good approximation for 𝑃(𝛿, 𝜖, 𝑚, 𝑇) can be given by the value of 𝑃̃ (𝛿, 𝜖,
𝑚−1−𝑙 . 注: Appendix 2 では、ある場合には P(δ, ε, m, T) に対するよい近似が P の値 (δ, ε, T) によって与えられることを示す。
0.47
). This is not far from the
). これはそれほど遠くない。
0.56
𝑚+1 𝑇 intuitive sense that in high probability, the parameter 𝜆 is close to the value 𝜆 =
𝑚+1 𝑇 高確率では、パラメータ λ は値 λ = に近いという直観的な感覚
0.53
𝑚 𝑇 . Assume now that we allow false alarm in a pixel up to probability 𝑞.
𝑚 𝑇 . 現在、ピクセル内の偽アラームを確率 q まで許容していると仮定する。
0.49
We note that in general, 𝑞 should be much smaller than the allowed probability for false alarm in the whole field of view.
一般に、q は視野全体における偽アラームの許容確率よりもはるかに小さいはずである。
0.61
Then, as 𝑄 is monotonically descending as function of 𝑁, there exists a minimal threshold 𝑁, such that 𝑄(𝛿, 𝜖, 𝑚, 𝑇, 𝑁) ≤ 𝑞.
このとき、Q は N の函数として単調に下降するので、Q(δ, ε, m, T, N) ≤ q となるような最小のしきい値 N が存在する。
0.81
For this specific threshold 𝑁 we get that the probability of a pixel with 𝑁 ≤ 𝑛 events in 𝐿𝑝𝑒𝑟𝑖𝑜𝑑𝑖𝑐 out of 𝑚 events in total, to falsely be considered as the desired periodic signal, is smaller than 𝑞.
この特定のしきい値 N に対して、合計 m 個の事象のうち、N ≤ n 個の事象を持つ画素の確率は、望まれる周期的信号として誤ったものとみなすことができ、q よりも小さい。
0.68
3
3
0.42
英語(論文から抽出)
日本語訳
スコア
Thus, in order to ensure that the signal in the pixel corresponds to the desired period with high enough probability, we need to check whether 𝑛 is at least 𝑁 or not.
したがって、画素内の信号が十分に高い確率で所望の周期に対応することを保証するためには、n が少なくとも N であるかどうかを確認する必要がある。
0.77
In the following lines, we explain how to answer this question without evaluating 𝑁 directly.
次の行では、Nを直接評価することなく、この質問に答える方法について説明する。
0.63
Observe first that we have 𝑄(𝛿, 𝜖, 𝑚, 𝑇, 𝑛) ≤ 𝑄(𝛿, 𝜖, 𝑚, 𝑇, 𝑁) ⇔ 𝑁 ≤ 𝑛.
However, as by definition, 𝑄(𝛿, 𝜖, 𝑚, 𝑇, 𝑁) is the maximal value of 𝑄 under the restriction 𝑄 ≤ 𝑞 for these specific parameters 𝛿, 𝜖, 𝑇, 𝑀, we actually get that in order to check whether 𝑁 ≤ 𝑛, all we need is to check the validity of the inequality
しかし、定義によれば、Q(δ, ε, m, T, N) はこれらの特定のパラメータ δ, ε, T, M に対する制限 Q ≤ q の下での Q の最大値であり、N ≤ n かどうかを確認するためには、不等式の有効性を確認する必要がある。
0.90
In other words, the value of 𝑄 = 𝑄(𝛿, 𝜖, 𝑚, 𝑇, 𝑛) measures if the signal satisfies the allowed probability for false alarm.
言い換えると、q = q(δ, ε, m, t, n) の値は、信号が誤警報の許容確率を満たすかどうかを測定する。
0.79
𝑄(𝛿, 𝜖, 𝑚, 𝑇, 𝑛) ≤ 𝑞.
𝑄(𝛿, 𝜖, 𝑚, 𝑇, 𝑛) ≤ 𝑞.
0.43
Experiment In general, the algorithm has better performance with detecting fast rising periodic signals, as in these cases one can choose the error parameter 𝜖 of the algorithm to be relatively small.
However, in order to feel its effectiveness lower bound, we demonstrate it here on the periodic signal of a streetlight powered by sinusoidal current of the electrical grid.
we need to choose 𝜖 to its highest reasonable value, namely 𝜖 =
ε を最も妥当な値、すなわち ε = に選ばなければならない。
0.85
𝛿 2 . In this case, as the
𝛿 2 . この場合は
0.42
frequency of the signal is 100 Hz, we have 𝛿 = 10 𝑚𝑠, and hence 𝜖 = 5 𝑚𝑠.
信号の周波数は 100 Hz であり、δ = 10 ms であり、ε = 5 ms である。
0.84
We note that by exploring the signal carefully, one might be able to give a better estimation for
注意深く信号を調べることで、より良い推定ができるかもしれないことに注意する。
0.75
the optimal 𝜖, but for the demonstration here we will content with 𝜖 =
最適 ε であるが、ここでのデモンストレーションは ε = で満足する。
0.83
𝛿 2 . Our experimental setup is built up of two cameras staring on approximately the same urban view: Prophesee Gen4 event-camera with resolution of 1M pixel, and a CMOS frame-camera for guidance.
Figure 1 shows the intersection of the two cameras field of view, seen by the frame camera.
図1は、フレームカメラで見る2つのカメラの視野の交差点を示しています。
0.82
This intersection covers the field of view of approximately 1000×600 pixels of the event camera.
この交差点は、イベントカメラの約1000×600ピクセルの視野をカバーする。
0.82
In Figure 1 we point out some well seen sun glittering coming from the buildings on the mountains, but there are also some intensive pixel size challenging sun glittering signals from the closer urban view that cannot be easily seen in the picture.
Now, in order to demonstrate how the algorithm distinguishes between the periodic signal of the streetlight, and all the other events, we recorded one minute of the scene, i.e. 𝑇 = 1 𝑠 .
さて、どのようにアルゴリズムが街灯の周期的信号と他の全てのイベントとを区別するかを示すために、我々はシーンの1分間、すなわち t = 1 s を記録しました。
0.82
Then, we applied the algorithm on the recording, and evaluated the function 𝑄(𝛿, 𝜖, 𝑚, 𝑇, 𝑛) for each pixel.
そして,このアルゴリズムを記録に適用し,各画素の関数Q(δ, ε, m, T, n)を評価した。
0.77
4
4
0.42
英語(論文から抽出)
日本語訳
スコア
Sun glittering Street light Figure 1: The intersection of the frame-camera and event-camera field of view.
太陽の輝き 街灯 図1: フレームカメラとイベントカメラの視野の交差点。
0.63
The logarithmic graph in Figure 2 shows the number of pixels in the intersection field of view that reported more than 5 events during the recording, for which the probability function 𝑄(𝛿, 𝜖, 𝑚, 𝑇, 𝑛) is smaller than the value of the allowed probability for false alarm given in the horizontal axis.
図2の対数グラフは、記録中に5つ以上の事象を報告したビューの交叉領域における画素数を示し、その確率関数 Q(δ, ε, m, T, n) は水平軸に与えられる偽アラームの許容確率の値よりも小さい。
0.77
In addition, it shows how many pixels out of them, are related to the periodic signal of the streetlight.
さらに、それらのうち何ピクセルが街灯の周期的な信号に関連しているかを示す。
0.73
Full field of view
フル・フィールド・オブ・ビュー
0.56
Street light 1.00E+04
街灯 1.00E+04
0.33
1.00E+03 1.00E+02
1.00E+03 1.00e+02
0.24
1.00E+01 1.00E+00
1.00E+01 1.00E+00
0.17
I S L E X P F O R E B M U N
私 s l e x p f o r e b m u n である。
0.69
1.E+00 1.E-02
1.E+00 1.E-02
0.20
1.E-04 1.E-06
1.E-04 1.e-06
0.28
1.E-08 1.E-10
1.e-08 1.E-10
0.37
ALLOWED PROBABILITY FOR FALSE ALARM
偽アルアームの追従確率
0.56
Figure 2: Number of pixels with probability function value smaller than the allowed probability for false alarm.
図2: 偽アラームの許容確率よりも、確率関数値の小さい画素数。
0.73
The graph shows that when the probability for false alarm is smaller than 10−5, we remain with only 5 pixels that have enough events in the output of the algorithm, and all of them are related to the streetlight.
of these 5 pixels is at least 2 orders of magnitude less than 10−5.
これらの5ピクセルのうち少なくとも2桁は10−5より小さい。
0.79
In other words, there is a very clear dichotomy between the probability function of these 5 pixels and the one of the other pixels.
言い換えれば、これらの5ピクセルの確率関数と他のピクセルの1つの間には、非常に明確な二分法がある。
0.70
Hence, these 5 pixels are very well distinguished by the algorithm and the analysis presented here.
したがって、この5ピクセルはアルゴリズムと解析によって非常によく区別される。
0.74
One can also see in Figure 2 that there were additional 9 more pixels that have reported more than 5 events during the recording, and their events are related to the streetlight, but were flickering more weakly, and hence the algorithm could not distinguish them effectively from other random signals.
In Table 1 we give the values of the variables that were involved in evaluating the probability function of all the pixels that responded to the flickering street light: the 5 distinguished pixels and the other 9 with weak signal, sorted according to the value of the probability function.
We remark that by analyzing the periodic signals more carefully, one can find a better value for the error parameter 𝜖, say 𝜖 = 3 𝑚𝑠, and be able to add pixels 6-8 in Table 1 to the list of the distinguished pixels.
周期的な信号をより慎重に分析することで、ε = 3 ms といった誤差パラメータ ε に対してより良い値を見つけ、卓1 に 6-8 の画素を区別されたピクセルのリストに追加することができる。
0.84
We also remark that obviously, analyzing a longer recording can help to detect these pixels as well.
98 ≤ 𝑛 ≤ 99 which is very close to the expected number
98 ≤ n ≤ 99 で、期待数に非常に近い。
0.81
𝑇 𝛿 = 100 , but 𝑚 varies
𝑇 𝛿 = 100 ですが m は異なります
0.55
between 160 ≤ 𝑚 ≤ 554.
between 160 ≤ 𝑚 ≤ 554.
0.42
The reason 𝑚 is much higher than the expected number of events in the pixel, is that in each of the periods, the rise of signal triggered more than one event: as powerful as the signal was, more events were triggered in each of the periods.
m がピクセル内の期待されるイベント数よりも大きい理由は、各周期において信号の上昇が複数のイベントをトリガーするからである。 訳抜け防止モード: m がピクセル内の期待されるイベント数よりはるかに高い理由。 それぞれの期間において、信号の上昇が複数のイベントをトリガーすること。 信号が強力であったように、各期間により多くの出来事が引き起こされた。
0.70
Now, focusing on the relation between 𝑚 and 𝑄, one can see that 𝑄 reached a higher value when 𝑚 = 342 (Pixel 5) than in the case 𝑚 = 160 (Pixel 3).
This phenomenon is due to the intuitive understanding, that tuning the parameter 𝑚 influences the probability function 𝑄 in two opposite directions.
この現象は、パラメータ m のチューニングが確率関数 Q に2つの反対方向に影響を与えるという直感的な理解が原因である。 訳抜け防止モード: この現象は直感的な理解が原因で パラメータ m のチューニングは2つの逆方向の確率関数 q に影響する。
0.86
From one hand, when 6
片手から、いつまで 6
0.48
英語(論文から抽出)
日本語訳
スコア
𝑚 grows it means that there are more events, and hence it should be easier to reach the threshold 𝑁, so 𝑄 should grow.
m が成長するということは、より多くのイベントがあることを意味するため、しきい値 n に到達しやすいため、q は大きくなるべきである。
0.69
On the other hand, when 𝑚 grows, the rate grows as well, and hence the chance for a specific random event to be far from the previous one in the right interval gets smaller, so 𝑃 decreases, and hence 𝑄 should decrease also.
This observation is summarized in the following proposition, showing that the function 𝑄 does encode these two opposite effects.
この観察は以下の命題で要約され、函数 Q がこれらの2つの反対効果を符号化していることを示す。
0.61
Proposition: Given a specific value for the variables 𝛿, 𝜖, 𝑇, 𝑛, with 𝜖 ≤
命題: ε ≤ の変数 δ, ε, T, n に対して特定の値を与える。
0.78
𝛿 2 and 𝑛 > 0,
𝛿 2 n > 0 である。
0.55
one has: In particular, 𝑄 reaches a maximum value as function of 𝑚.
一つは 特に、Q は m の関数として最大値に達する。
0.73
Proof: Writing 𝑃 = 𝑃(𝛿, 𝜖, 𝑚, 𝑇) and using the assumption 𝑛 > 0, one has
証明: P = P(δ, ε, m, T) を書き、n > 0 の仮定を用いる。
0.71
lim 𝑚→0 𝑄 = lim 𝑚→∞
lim m→0 Q = lim m→∞
0.39
𝑄 = 0. 𝑄(𝛿, 𝜖, 𝑚, 𝑇, 𝑛) = ∑ (
𝑄 = 0. 𝑄(𝛿, 𝜖, 𝑚, 𝑇, 𝑛) = ∑ (
0.43
𝑚−1 𝑚 − 1 𝑙
𝑚−1 𝑚 − 1 𝑙
0.38
) 𝑃𝑙(1 − 𝑃)𝑚−1−𝑙
) 𝑃𝑙(1 − 𝑃)𝑚−1−𝑙
0.41
𝑙=𝑛 𝑚−1 = 𝑃 ⋅ ∑ (
𝑙=𝑛 𝑚−1 = 𝑃 ⋅ ∑ (
0.34
𝑙=𝑛 𝑚−2 = 𝑃 ⋅ ∑
𝑙=𝑛 𝑚−2 = 𝑃 ⋅ ∑
0.34
𝑙=𝑛−1 𝑚−2 𝑚 − 1
𝑙=𝑛−1 𝑚−2 𝑚 − 1
0.31
𝑙 ) 𝑃𝑙−1(1 − 𝑃)𝑚−1−𝑙
𝑙 ) 𝑃𝑙−1(1 − 𝑃)𝑚−1−𝑙
0.39
𝑚 − 1 𝑙 + 1
𝑚 − 1 𝑙 + 1
0.43
⋅ ( 𝑚 − 2 𝑙
⋅ ( 𝑚 − 2 𝑙
0.43
) 𝑃𝑙(1 − 𝑃)𝑚−𝑙−2 =
) 𝑃𝑙(1 − 𝑃)𝑚−𝑙−2 =
0.42
≤ 𝑃 ⋅ 𝑚 ⋅ ∑ (
≤ 𝑃 ⋅ 𝑚 ⋅ ∑ (
0.43
𝑙=𝑛−1 𝑚 − 2
𝑙=𝑛−1 𝑚 − 2
0.31
𝑙 ) 𝑃𝑙(1 − 𝑃)𝑚−𝑙−2
𝑙 ) 𝑃𝑙(1 − 𝑃)𝑚−𝑙−2
0.42
= 𝑃 ⋅ 𝑚 ⋅ 𝑄(𝛿, 𝜖, 𝑚 − 1, 𝑇, 𝑛 − 1) ≤ 𝑃 ⋅ 𝑚 1
= 𝑃 ⋅ 𝑚 ⋅ 𝑄(𝛿, 𝜖, 𝑚 − 1, 𝑇, 𝑛 − 1) ≤ 𝑃 ⋅ 𝑚 1
0.43
1 = 𝑚 ⋅ 𝑇𝑚+1 (
1 = 𝑚 ⋅ 𝑇𝑚+1 (
0.43
(𝑇 + 𝛿 − 𝜖)𝑚+1 −
(𝑇 + 𝛿 − 𝜖)𝑚+1 −
0.49
(𝑇 + 𝛿 + 𝜖)𝑚+1)
(𝑇 + 𝛿 + 𝜖)𝑚+1)
0.49
𝑚→0 → 0.
𝑚→0 → 0.
0.42
As we assume that 𝜖 ≤
ε ≤ と仮定すると
0.70
𝛿 2 , we have 𝛿 − 𝜖 > 0, and hence we also have
𝛿 2 δ − ε > 0 であり、従って δ − ε > 0 である。
0.51
𝑄(𝛿, 𝜖, 𝑚, 𝑇, 𝑛) ≤
𝑄(𝛿, 𝜖, 𝑚, 𝑇, 𝑛) ≤
0.42
𝑚 ⋅ 𝑇𝑚+1 (𝑇 + 𝛿 − 𝜖)𝑚+1 =
𝑚 ⋅ 𝑇𝑚+1 (𝑇 + 𝛿 − 𝜖)𝑚+1 =
0.44
𝑚 (1 + 𝛿 − 𝜖
𝑚 (1 + 𝛿 − 𝜖
0.43
𝑇 ) 𝑚→∞ → 0
𝑇 ) 𝑚→∞ → 0
0.41
𝑚+1 As required.
𝑚+1 必要に応じて
0.51
Appendix 2 We want to show here that 𝑃̃ (𝛿, 𝜖, when 𝑚 is not too big and 𝛿 ≪ 𝑇 .
付録2 ここで m が大きすぎないとき、p は (δ, ε) であり、δ は t であることを示す。
0.67
Using the aforementioned assumption, we can expand the expressions for 𝑃 and 𝑃̃ as shown below.
上記の仮定を用いて、以下のように P と P の式を拡張することができる。
0.73
) gives good approximation to 𝑃(𝛿, 𝜖, 𝑚, 𝑇)
) P(δ, ε, m, T) によい近似を与える
0.83
𝑇 𝑚+1 This computation shows that up to the second order of the expansions, the difference between the expressions is big as (𝑚 + 1) 2𝜖𝛿 𝑇2 , which is small under the aforementioned assumptions.
1. A. Censi, J. Strubel, C. Brandli, T. Delbruck and D. Scaramuzza, "Low-latency localization by Active LED Markers tracking using a Dynamic Vision Sensor," 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp.
1. A. Censi, J. Strubel, C. Brandli, T. Delbruck, D. Scaramuzza, “Low-latency Localization by Active LED Markers tracking using a Dynamic Vision Sensor”. IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 2013 訳抜け防止モード: 1. A. Censi, J. Strubel, C. Brandli, T. Delbruck D. Scaramuzza氏は,“ダイナミックビジョンセンサを用いたアクティブLEDマーカトラッキングによる低レイテンシローカライゼーション”について述べている。 2013 IEEE / RSJ International Conference on Intelligent Robots and Systems,pp
0.60
891-898, 2013.
891-898, 2013.
0.42
2. G. Gallego, T. Delbruck, G. Orchard, C. Bartolozzi, B. Taba, A. Censi, S. Leutenegger, A. J. Davison, J. Conradt, K. Daniilidis and D. Scaramuzza, "Eventbased Vision: A Survey," IEEE transactions on pattern analysis and machine intelligence, vol.
2. G. Gallego, T. Delbruck, G. Orchard, C. Bartolozzi, B. Taba, A. Censi, S. Leutenegger, A. J. Davison, J. Conradt, K. Daniilidis, D. Scaramuzza, "Eventbased Vision: A Survey", IEEEのパターン分析とマシンインテリジェンスに関するトランザクション。
0.69
44, no. 1, pp. 154-180, 2020.
44, No. 1, pp. 154-180, 2020。
0.48
3. Q. Sabatier, S.
3. Q. Sabatier, S。
0.45
-H. Ieng and R. Benosman, "Asynchronous Event-Based Fourier Analysis," IEEE Transactions on Image Processing, vol.
-h。 ieng と r. benosman, "asynchronous event-based fourier analysis", ieee transactions on image processing, vol.2 (英語)
0.70
26, no. 5, pp. 2192-2202, 2017.
26 no. 5, pp. 2192-2202, 2017年。
0.89
4. M. H. Tayarani-Najaran and M. Schmuker, "Event-Based Sensing and Signal Processing in the Visual, Auditory, and Olfactory Domain: A Review," Frontiers in Neural Circuits, vol.
4. M. H. Tayarani-Najaran, M. Schmuker, "Event-based Sensing and Signal Processing in the Visual, Auditory and Olfactory Domain: A Review", Frontiers in Neural Circuits, Vol. 訳抜け防止モード: 4. m. h. tayarani - najaran and m. schmuker, "イベントに基づくセンシング" 視覚・聴覚・嗅領域における信号処理 : レビュー」 神経回路におけるフロンティア
0.54
15, 2021.
15, 2021.
0.42
David El-Chai Ben-Ezra, Remote Sensing Department, Soreq NRC, Yavne, Israel 81800 (davidbe@soreq.gov.i l)
David El-Chai Ben-Ezra, Remote Sensing Department, Soreq NRC, Yavne, Israel 81800 (davidbe@soreq.gov.i l)
0.43
Ron Arad, Remote Sensing Department, Soreq NRC, Yavne, Israel 81800 (fnarad@soreq.gov.il )
Ron Arad, Remote Sensing Department, Soreq NRC, Yavne, Israel 81800 (fnarad@soreq.gov.il )