SYSTEM FOR MONITORING AN ACOUSTIC SCENE OUTSIDE A VEHICLE

Abstract

A system for monitoring an acoustic scene outside a vehicle; the system including: a vehicle with wheels and a trunk, an acoustic sensor disposed in the trunk, a control unit operatively connected to the acoustic sensor, and at least one neural network operatively connected to the control unit, and trained in such a way to correlate the characteristics of an audio signal with types of road surfaces; the control unit is configured in such a way to receive an audio signal detected by the acoustic sensor while the vehicle is traveling, extract the characteristics of the audio signal and input said characteristics of the audio signal to the neural network in order to identify the type of road surface covered by the vehicle wheels.

Claims

1. System for monitoring an acoustic scene outside a vehicle; said system comprising: a vehicle with wheels, a trunk and a passenger compartment that is distinct and separate from the trunk; an acoustic sensor disposed in said trunk; a control unit operatively connected to the acoustic sensor; and at least one neural network operatively connected to the control unit, which is suitably trained in order to correlate the characteristics of an audio signal with types of road surfaces; wherein said control unit is configured in such a way to receive an audio signal detected by said acoustic sensor while the vehicle is traveling, extract the characteristics of said audio signal and provide said characteristics of said audio signal to the neural network in order to identify the type of road surface covered by the vehicle wheels; characterized in that said trunk comprises a bottom made of metallic material that separates the trunk from the exterior and a partition wall disposed in the trunk, in such a way to define a lower compartment; wherein said acoustic sensor is disposed in said lower compartment of the trunk.

2. The system of claim 1, wherein said acoustic sensor is disposed on said bottom made of metallic material of the trunk.

3. The system of claim 1, wherein said neural network comprises a Convolutional Neural Network (CNN) that is trained to identify the road surface conditions based on the signal.

4. The system of claim 1, wherein said characteristics of the audio signal extracted by the control unit are Auditory Spectral Features obtained from the Mel spectrograms of the audio signal.

Description

BRIEF DESCRIPTION OF THE SEVERAL VIEWS OF THE DRAWINGS

[0045] The system for monitoring an acoustic scene outside a vehicle according to the invention is defined in claim 1.

[0046] Additional features of the invention will be clear from the following detailed description, which refers to a merely illustrative, not limiting embodiment, as shown in the appended figures, wherein:

[0047] FIG. 1 is a chart that illustrates the difference between the spectrum of an audio signal acquired by an acoustic sensor provided with a dry anti-wind sponge and the one acquired by an acoustic sensor provided with a wet anti-wind sponge;

[0048] FIG. 2 is a diagrammatic top view of a vehicle wherein a plurality of mounting positions of the sensors is indicated;

[0049] FIG. 3 is a diagrammatic bottom view of the vehicle of FIG. 2, which shows four positions in the proximity of the wheels of the vehicle and one position in the trunk with a broken line;

[0050] FIG. 4 is a chart that shows a signal attenuation as a function of the frequency, between the driving position and an acoustic sensor disposed in the trunk;

[0051] FIG. 5 is a partial sectional view of a trunk of the vehicle of FIG. 2;

[0052] FIG. 6 is a block diagram that diagrammatically shows the operation of the system according to the invention.

DETAILED DESCRIPTION OF THE INVENTION

[0053] With reference to FIGS. 2 to 6, the acoustic sensor system for automotive application according to the invention is disclosed, which is generally indicated with reference numeral (100).

[0054] FIGS. 2 and 3 illustrate a vehicle (1). The vehicle (1) comprises wheels (6), a trunk (3) and a passenger compartment that is distinct and separate from the trunk (3).

[0055] A plurality of mounting positions for the acoustic sensors (microphones) is identified in the vehicle (1), in order to detect an acoustic scene outside the vehicle, such as for example the road surface condition (dry, wet, smooth, rough, etc.).

[0056] According to the prior art, the following eight mounting positions are generally used for the acoustic sensors: [0057] a position (E) in the engine compartment, [0058] a position (FL) in front of the left front wheel, [0059] a position (FR) in front of the right front wheel, [0060] a position (RL) in front of the left rear wheel, [0061] a position (RR) in front of the right rear wheel, [0062] a position (DP) in the license plate holder compartment, [0063] a position (ID) in the passenger compartment above the head of the driver, and [0064] a position (IB) in the passenger compartment above the passenger who is sitting in the back in the most distant position from the driver.

[0065] The present invention has examined the aforesaid eight positions and has demonstrated that they are not convenient; therefore, the present invention proposes a new mounting position for an acoustic sensor (2) in the trunk (3).

[0066] With reference to FIG. 5, the trunk (3) generally comprises a bottom (30) made of metallic material that separates the trunk (3) from the exterior. A partition wall (31) is disposed in the trunk in such a way to define a lower compartment (32) between the bottom (30) and the partition wall (31). The lower compartment (32) is generally used to house a spare wheel or tools.

[0067] The system according to the invention provides for installing an acoustic sensor (2) in the position (T) inside the trunk (3), in particular in the lower compartment (32) used for housing the spare wheel that is situated under the partition wall (31) of the trunk. Preferably, the acoustic sensor (2) is positioned on the bottom (30) made of metallic material of the trunk.

[0068] The inventive solution makes the acoustic sensor (2) free from the interference caused by the weather agents because the acoustic sensor (2) is situated inside the vehicle. Therefore the acoustic sensor (2) does not need to be protected with anti-wind protections or waterproofing systems, with evident advantages both acoustically and economically.

[0069] Moreover, the acoustic sensor (2) is protected against the interference that is typical of the acoustic sensors disposed inside the passenger compartment (such as for example acoustic sensors disposed in the positions (ID) and (IB)) because the presence of the back seats of the vehicle, of the rear shelf and of the partition wall (31) of the trunk attenuates the sounds coming from the passenger compartment, acoustically separating the two spaces (passenger compartment and lower compartment (32) under the partition wall (31) of the trunk).

[0070] FIG. 4 illustrates a chart that shows the isolation (signal attenuation) measured between an acoustic sensor disposed in the position (ID) above the head of the driver and the acoustic sensor (2) disposed in the position (T) in the trunk. Such a chart refers to a Mercedes Class A vehicle. As shown in FIG. 4, the attenuation increases when the frequency increases. In particular, attenuation values higher than 20 dB are obtained for frequencies around 1 KHz.

[0071] Moreover, it must be noted that the issue of acoustic isolation from the exterior is considerably mitigated because the acoustic sensor (2) is disposed in adjacent position to the bottom wall (30) made of metallic material of the trunk, which is provided with a lower soundproofing than the body of the passenger compartment.

[0072] Additionally, the system according to the invention has some practical advantages. For example, in case of a minor accident or a bumper-to-bumper crash, the acoustic sensor (2) is protected against damage because it is not mounted inside the bumper of the vehicle.

[0073] Moreover, it is not necessary to modify the pre-existing spaces, which can maintain their external appearance, because the acoustic sensor (2) is concealed during the daily use of the vehicle.

[0074] As it is known, a vehicle is provided with rear lamps that are electrically powered. Therefore, the electrical connections of the rear lamps can be used for powering the acoustic sensor (2). Obviously, when designing the trunk, a specific housing should be provided for the acoustic sensor (2) in order to facilitate the electrical connections and the installation of the acoustic sensor.

[0075] The system according to the invention remedies the aforementioned practical issues and guarantees excellent results in the detection of the road surface because the acoustic sensor (2) is proximal to the road surface and the aforementioned interference is removed.

[0076] With reference to FIG. 6, the acoustic sensor (2) is electrically connected to a control unit (4) of the vehicle. The control unit (4) is connected to a neural network (5) that is suitably trained and inputted with features obtained from the audio signal detected by the acoustic sensor (2). The neural network (5) has a configuration and a set of weights.

[0077] The term configuration indicates the combination of layers and the number and interconnection of the neurons of the neural network. The neuron is a structure that is provided with inputs and produces an output using a special activation function that varies according to the type of neuron. The weights are parameters that are used to weigh the amount in which a specific input of a neuron affects its output; otherwise said, the weights are used to manage the interconnections between neurons

[0078] As such, the neural network can be considered as a parametric circuit, wherein the architecture and the quantity of parameters can be established in advanced.

[0079] The optimal weights and the optimal configuration are obtained during a first step, which is defined as learning, and determine the understanding of the problem by the neural network.

[0080] After being optimally trained, the neural network (5) is inputted with data of the audio signal from the acoustic sensor (2). By processing the audio signal from the acoustic sensor (2), the neural network (5) returns information on the road surface to the control unit (4).

[0081] During a testing step of the neural network (i.e. when an algorithm of the neural network operates without any knowledge support from the exteriorthe operating condition of interest), the audio signal is pre-processed by a DSP algorithm to extract the characteristic features of the signal in the time-frequency domain for each recording that is analyzed. These features are passed to the neural network that outputs the type of road surface (or the wet/dry condition) relative to the recording.

[0082] The neural network is a parametric circuit, wherein the parameters have been determined by means of the so-called training step, which uses a suitable learning algorithm that uses the knowledge of the features of the incoming audio signal and of the relative class that represents the type of road surface (information available after a manual labeling action on a certain number of recordingstraining dataset). Obviously, the recordings of the training step and of the testing step are different.

[0083] For illustrative purposes, the neural network (5) can be a Bidirectional Long Short Term Memory (BLSTM). The features that are extracted from the audio signal detected by the acoustic sensor (2) can be Auditory Spectral Features (ASF), i.e. features obtained from Mel spectrograms of the audio signal.

[0084] Alternatively, an efficaciously trained Convolutional Neural Network (CNN) capable of distinguishing the type of road surface (smooth/rough) can be used.

[0085] Obviously, the neural network (5) provides for training on a large dataset that comprises types of different tires (summer tires (S) and winter tires (W)) and makes a comparison on the real generalization and transfer learning capacities of the system, comparing the potential on both types of tires.

[0086] The system performance is evaluated by using metrics of statistic type. Specifically, an F1-score metrics is used to measure the accuracy of a binary classification by means of a harmonic mean of precision and recall, i.e. using the knowledge of correct and incorrect predictions for each class. In such a case, being a binary classification, the classes can be considered as positive and negative, defining [0087] TP: correct positive prediction, [0088] TN: correct negative prediction, [0089] TP: incorrect positive prediction, [0090] FN: incorrect negative prediction.

[0091] Therefore, the metrics can be formulated as follows:

[00001]$F.Math..Math.1-\mathrm{score}.Math.=\frac{2.Math.\mathrm{TP}}{2.Math.\mathrm{TP}+\mathrm{FP}+\mathrm{FN}}$

[0092] Following are the results obtained for the classification of dry/wet and smooth/rough road surface to show the performance that can be obtained according to the position of the acoustic sensors and to the type of tires (winter tires (W) or summer tires (S)).

TABLE-US-00001 TABLE 1 Mic TRAIN/TEST F1-score DP W/W 98% DP W/S 1% DP S/W 9% DP S/S 92% ID W/W 67% ID W/S 74% ID S/W 62% ID S/S 96% IB W/W 99% IB W/S 94% IB S/W 96% IB S/S 98% T W/W 96% T W/S 95% T S/W 97% T S/S 99%

[0093] Table 1 shows the results related with the recognition of the road surface in terms of dry/wet in four acoustic sensors disposed in the four best positions illustrated in FIG. 2. It has been demonstrated that the best positions are the positions (DP, IB, ID and T) in the recognition of the wet road surface using different datasets and CNNs. As shown in Table 1, in three cases out of four, the best result is obtained with the acoustic sensor disposed in the position (T) in the trunk.

TABLE-US-00002 TABLE 2 Mic TRAIN/TEST F1-score DP S/S 92.1% DP W/W 85.6% DP S/W 83.2% DP W/S 82.9% ID S/S 93.0% ID W/W 89.3% ID S/W 76.7% ID W/S 87.5% IB S/S 92.7% IB W/W 90.8% IB S/W 72.3% IB W/S 84.4% T S/S 93.2% T W/W 91.8% T S/W 88.7% T W/S 88.7% RR S/S 92.7% RR W/W 87.5% RR S/W 81.8% RR W/S 80.3%

[0094] Table 2 shows the results related with the recognition of the smooth/rough road surface using the same acoustic sensors used for Table 1, with the addition of one acoustic sensor disposed in the position (RR) in front of the right rear wheel, which can be used in this case because the signals to be classified are acquired with dry road surface. Also in this case, the tests are made using an algorithm that uses the CNNs. The results are referred to different types of datasets recorded using two different types of tires, i.e. summer tires (S) and winter tires (W).

[0095] The training of the neural networks is made on a type of tires, whereas the test (of which the F1-score is shown) can be made on the same type of tire or on the other type of tire. The purpose of such an assessment is to assess whether and how much the different type of tire affects the final result.

[0096] As shown in Table 2, in all four cases, the best result is obtained with the acoustic sensor disposed in the position (T) in the trunk.

[0097] The results of Table 1 and Table 2 show that the performance of both systems is acceptable and fully meets the generalization requirement; however, evidently, the performance is heavily affected by the position of the acoustic sensor and by the presence of the aforementioned interference. Evidently, the position of the acoustic sensor is a fundamental aspect in the examples given and, more generally, in all applications that use acoustic sensors for monitoring the acoustic scene outside the vehicle.

[0098] Finally, it must be noted that the solution of the present invention seems to be the best solution possible among the solutions that are studied in the majority of the cases and is considerably different from the other solutions. Therefore, the results are an important experimental confirmation of the aforementioned advantages.

[0099] Numerous equivalent variations and modifications, which are within the reach of an expert of the field and fall in any case within the scope of the invention as disclosed by the appended claims, can be made to the present embodiment of the invention.