CASCADE OF CLASSIFIERS

Technical Field

The present disclosure relates to a method detects and classifies a shock and muzzle signal from acoustic system using a cascade of classifiers.

Background

Existing methods only use spectrogram based neural network approaches that take long processing time since spectrogram is computed in sliding window fashion and neural network is computationally costly for evaluation at every sliding window. Also, proposed framework rejects shocks with simple feature engineering, and eliminates echo in urban environment with correlation filtering. Available shock and muzzle detectors do not scale with the increasing number of microphones. Also, existing methods create excessive number of candidates which might create false shots.

Existing methods do not detect and classify shock and muzzle signals under the same framework considering the real-time constraints for low power processors. The patent numbered EP2204665 is an example for above stated drawbacks. The patent document proposes a method for identifying a muzzle blast within a signal generated by an array of acoustic sensors forming an antenna comprises defining a width for a time window corresponding to a time required for the muzzle blast to traverse the array of acoustic sensors and detecting a shockwave within the generated signal. After detecting the shockwave, total energy of the generated signal in a series of consecutive time windows as a function of time and as a function of the number of acoustic sensors that generated the signal is measured, each time window having the defined width. One of the consecutive time windows having a measured total energy greater than each of the remainder of the consecutive time windows is identified; and the identified time window is associated as corresponding to the muzzle blast.

Summary

Proposed method detects and classifies a shock and muzzle signal from acoustic system. The method works with both single and multiple microphones, such that when formed an acoustic array a time-difference-of-arrival (TDOA) reveals the shooter angle. Time difference between shock and muzzle can be used for range estimation.

The invention uses a cascade of classifiers with increasing complexity to eliminate false signals, such that computational complexity remains low and enables real-time performance for an array of microphones.

On the other hand, the inventive technique eliminates false shock and muzzles with feature engineering at early stages, and then with advanced techniques for rare potential candidates at later stages. Classified muzzles are further cross correlated for echo elimination for operation in urban environment.

The present method eliminates false signals with classifiers of increasing complexity such that only candidates are evaluated in complex and accurate classifiers, and proposed algorithm scales with increasing number of microphones.

Proposed algorithm verifies the validity of a shock using simple peak detector. Classified muzzles are further examined in terms of correlation. Low Pearson correlation muzzles stem from echo in urban environment. Such a technique allows system to operate in highly echoic environment.

Brief Description of the Drawings

Figure 1 shows spectrogram, raw data and FFT of shock and muzzle signals of a shot.

Figure 2 shows filtering and classification process.

Figure 3 shows the flowchart of detection and classification of shock signals.

Figure 4 shows the flowchart of detection and classification of muzzle signals.

Figure 5 shows shock/muzzle selection scheme.

Figure 6 shows normalized amplitude and time graphic of a shock signal including Tsize demonstration between peak points.

Figure 7 shows spectrogram of power normalized signals reveal the type of signal.

Detailed Description

Gunshot acoustics are mainly composed of a shock wave and muzzle blast. A supersonic bullet generates a shock wave pattern. Shock wave has a conic pattern which deteriorates as bullet travels. A sample of both waveforms are given in Figure 1. In figure, shock and muzzle belong to the same shot. It is obvious that across figure both shock and muzzle vary drastically. In first row of Figure 1 , spectrogram is given. Second row of figure show raw data, while third row show the Fast Fourier Transform (FFT).

As shown in Figure 2, in the method, a processor records the acoustic signals from the microphone to a database. Database contains shock, muzzles and noise signals. Acoustic signals are FIR (Finite Impulse Response) / HR (Infinite Impulse Response) filtered using a high pass and band pass filters. High pass filter aims to extract shock candidates at original scale without any decimation. Band bass filter aims to extract muzzles using decimation since muzzles span a larger time scale. Shock and muzzle candidates are stored on different databases and processed differently. Figure 3 and 4 demonstrates the flowcharts of shock and muzzle classification independently.

Shock candidates are applied following operations consecutively:

• Shock candidate signal is preprocessed using RMS (Root Mean Square) normalization.

• Peak time difference (Tsize) computation that finds the time elapsed between the first positive peak and the first negative peak. (Figure 6)

• Logistic regression classifier is applied using FFT of the signal.

• Convolutional neural network is applied using spectrogram of the signal.

Muzzle candidate signal is applied following cascade operations to be detected as a muzzle:

• Muzzle candidate signal is preprocessed using RMS normalization.

• Decimation and band pass filtering in original size is operated using HR or FIR.

(implemented as LP + Decimate + HP)

• Logistic regression classifier is applied using FFT of the signal

• Convolutional neural network is applied using spectrogram of the signal

Echo of muzzles within urban environment creates a muzzle candidate that fulfills all muzzle criteria, so muzzles are aligned and the Pearson correlation is computed. Only muzzles with high correlation scores are reported in urban environment to eliminate echo conditions.

Figure 5 shows shock/muzzle selection scheme. All signals acquired by each microphone is filtered to get regions of interest which might contain shock and muzzles. Shock regions are extracted using FIR/IIR filtering with high pass filter, while muzzle regions are extracted using IIR/FIR filtering using band pass filter. Muzzle filtering also utilizes decimation factor (e.g. decimation factor of 16 since muzzle spans a time of 6 ms, while shock spans only 1.25 ms in average). Each incoming signal is power RMS normalized as following:

Then signal is exposed to a three-stage cascade classifier for efficiency. Signals whose IIR/FIR response small then a limit value are considered noise and rejected. Signals passing this level are FFT taken in a window (e.g. limit value is 5 and 512 point FFT is used), then classified using Logistic Regression (LR) to decide if a signal is noise or shock/muzzle. Finally, if a candidate passes previous levels, it is classified as shock or muzzle in final stage using spectrogram features and Convolutional Neural Network (CNN).

There can be multiple ripples in shock region that can correspond to a real shock signal. It is critical to select the first positive peak and the second negative peak of shock, since this is a critical feature for range estimation. A derivate based approach is proposed to locate these peaks. A five-point derivative is used for reliability, e.g. using coefficients of [1 -8 0 8 -1] / 12. Then, second order derivative is taken only using two points. A signal peak is where first order derivate switches signs, and second order derivative has a certain amplitude. This simple criterion allows peak time difference estimation with sample accuracy.

Given spectrogram of power normalized signals reveal the type of signal clearly as shown in Figure 7, which type can be noise, shock or muzzle. Although spectrogram computation is costly with respect to FFT, this approach yields a rich information. Combined with convolutional neural networks (CNN), this approach provides classification accuracy above %99, compared to logistic regression (LR) %89.