$\mathcal{A}$$\mathcal{C}$lassi$\mathcal{H}$onk: A System Framework to Annotate and Classify Vehicular Honk from Road Traffic

Noise level characterization is mainly done by monitoring, forecasting, and identifying the source of noise in the environment. Several techniques (like deploying heterogeneous sensors in different places of cities/smartphone-based monitoring) are adopted to measure the pollution level, where smartphone-based sensing is most popular one. Smartphones are basically used for spatial data collection, while fixed sensor-based devices are deployed for temporal data collection.

Smartphone-base monitoring: In Zipf \BOthers. (\APACyear2020); Jezdović \BOthers. (\APACyear2021), the author collected the data using smartphones and notified the efficacious urban planning corresponding to noise pollution exposure. Spatio-temporal patterns of noise pollution are also measured using smartphones in Zamora \BOthers. (\APACyear2017). Apart from this, the researchers in Maity, Polapragada\BCBL \BOthers. (\APACyear2022); Allen \BOthers. (\APACyear2009) have addressed a correlation between noise and air pollution to recognize the primary sources of environmental noise clearly. Due to the sensitivity issue of the smartphone and the lack of proper calibration techniques, infrastructure-based sensing seems more realistic.

Fixed sensor-based monitoring: To get the temporal variation of data, different calibrated sensors can be deployed throughout the city (mainly in the road traffic area) to measure the noise pollution level. In previous studies, the researchers developed a wireless sensor network Santini \BOthers. (\APACyear2008) in order to monitor noise pollution. However, nowadays, many advanced sensors are developed that can capture data with a high level of accuracy Bello \BOthers. (\APACyear2019). Raspberry Pi-controlled cloud-based sensor is also designed in Saha \BOthers. (\APACyear2018) to do the same. Moreover, an Arduino controller with IoT technology Ezhilarasi \BOthers. (\APACyear2017) is also used as a fixed sensing strategy to get the noise data from the environment. Like Santini \BOthers. (\APACyear2008), an advanced wireless sensor-based system is designed for the same in Segura-Garcia \BOthers. (\APACyear2014). Besides the advancement of technology, some pros exist that make the sensors unreliable. The sensors may provide erroneous data or stop functioning due to environmental hazards such as storms, rainfall, etc.

Noise level forecasting: To predict the noise pollution level in road traffic, the researchers Medina-Salgado \BOthers. (\APACyear2022) have developed several models and techniques. In Garg \BOthers. (\APACyear2015); Guarnaccia \BOthers. (\APACyear2017), the authors proposed ARIMA model-based time series traffic noise prediction model. Deep learning-based models are also used in several kinds of research to forecast pollution levels. In Navarro \BOthers. (\APACyear2020), the authors developed LSTM-based models to predict sound pollution levels in urban road traffic. In our previous work Maity, Trinath\BCBL \BOthers. (\APACyear2022), we also estimated the noise pollution level in the road traffic area by predicting the vehicular honk count using the E-LSTM model.

Noise source identification: Apart from the above mentioned techniques, noise pollution measurement by identifying the source is another way to monitor pollution. In Vera-Diaz \BOthers. (\APACyear2018); Suvorov \BOthers. (\APACyear2018), the author identified the source of sound using a deep neural network. Sound source identification using a microphone array configuration is another effective technique to measure pollution exposure. In Grondin \BBA Michaud (\APACyear2019), the author used an open and closed microphone array for sound localization. Wireless Acoustic Sensor Networks-based sound source identification is addressed in Cobos \BOthers. (\APACyear2017) by the authors.

The primary source of noise pollution is traffic noise, which is mainly comprised of different vehicular honks. Therefore, vehicular honk identification and classification of the different types of vehicle honk are essential to characterize the noise pollution level. Several researchers are already identified vehicular honks by adopting different strategies. These strategies are mainly classified into three categories: a) classifying environmental sound, where vehicular honk is one of the classes, b) determining the honks directly from the audio samples by using FFT, Spectrogram, or MFCC, and c) applications based on the honk signature. Details of the honk identification and sound classification methods are summarized in Table 2.

Environmental sound classification: There are numerous works that identify vehicular honking by classifying environmental sounds, where car honking is one of the class Piczak (\APACyear2015); Salamon \BBA Bello (\APACyear2017); Zhou \BOthers. (\APACyear2017); Khamparia \BOthers. (\APACyear2019); Abdoli \BOthers. (\APACyear2019); Mesaros \BOthers. (\APACyear2019); Chen \BOthers. (\APACyear2019); Demir \BOthers. (\APACyear2020); Ahmed \BOthers. (\APACyear2020); Mushtaq \BBA Su (\APACyear2020); Guzhov \BOthers. (\APACyear2021); Mu \BOthers. (\APACyear2021). Authors mainly used ESC-10, ESC-50, or UrbanSound8K dataset for sound classification. All the data are collected in a controlled environment where ambient sound is not present. Most of the cases, they modified the CNN models to increase the accuracy. The highest level of accuracy that they have achieved till yet is 97% for the ESC-10 dataset in Guzhov \BOthers. (\APACyear2021). However, the ESC-10 data set does not contain the honk class. Apart from ESC-10, the maximum accuracy achieved by the DCASE-2017 ASC dataset Demir \BOthers. (\APACyear2020) is 96.23%, which contains a car honk as a class.

Honk from raw samples and its applications: In Sen \BOthers. (\APACyear2010), the author used band-pass filtering to remove the noises from the raw audio samples and then applied FFT to detect the honks. They also calculated the duration of each honk in this work. Nevertheless, their selected threshold values for band-pass filters don’t guarantee a definite honk signature all the time. Furthermore, any framework/system was not designed that can identify honk automatically while moving. A similar kind of FFT-based honk detection technique is found in Dim \BOthers. (\APACyear2020). As an application of honk, the authors developed a system that assists the hearing-impaired person in driving. In another work Takeuchi \BOthers. (\APACyear2014), a smartphone-based system is implemented to detect honking and then generate alarm sounds for hearing-impaired people. Apart from these, an embedded system is designed to identify emergency honk Palecek \BBA Cerny (\APACyear2016), and MFCC-based honk detection in Banerjee \BBA Sinha (\APACyear2012). In our previous work Maity, Alim\BCBL \BOthers. (\APACyear2022), we have identified vehicular honk from raw audio samples in the presence of several ambient noises. Apart from this, we have also determined different features of the honk, like duration of the honk, the inter-honk gap between two successive honks, and honk count, etc., and based on these features, we have tried to identify the context of a location and recommend a honk-aware route, which is a healthier route in comparison to the google recommended route.

As per the works mentioned above, it is apparent that the researchers are working on honk identification and environmental sound classification techniques. However, fine-grained vehicular honk classification in terms of vehicle types is our primary objective of the work presented in this paper, and it is not yet done.

In this section, different vehicle honking features are analyzed by considering spatio-temporal characteristics. We observed 5 minutes of data for three different places (residential area, marketplace, high traffic) and verified the count of honks, SPL level, and correlation between them for each type of vehicle. Moreover, we performed manual data labeling for this experiment.

A spectrogram is a visual representation of a raw audio file, depicting the loudness of a signal over time at different frequencies present in a waveform. Similar to our previous work Maity, Alim\BCBL \BOthers. (\APACyear2022), we used spectrogram to determine honks from raw audio samples. To generate a spectrogram image, we perform Fourier Transformation, which decomposes the signals into frequencies and shows the amplitude of each frequency over time. Like Maity, Alim\BCBL \BOthers. (\APACyear2022), the current work generates spectrograms by dividing each audio sample into a 1-seconds segment, with the X-axis representing the time(sec), the Y-axis denoting frequency(Hz), and the amplitude of each frequency is represented by the different colors. Brighter color indicates a higher amplitude. As an example, we have considered a random 10-sec audio clip and converted it into spectrogram images, which is shown in Fig. 6. The figure shows that there are five possible honks (2 LWV, 1 MWV, and 2 HWV) within this 10-sec duration. Furthermore, Fig. 6, 6, and 6 illustrates the individual spectrogram image of audio data collected from LWV, MWV, and HWV environment for better understanding.

Deep learning classifiers require a significant volume of good-quality labeled data. In the absence of such a rich dataset related to the problem presented in this paper, we initially considered manual labeling. Nevertheless, manual labeling is simple, but it is labor-intensive, time-consuming, and more prone to human error. To deal with that, we employed two personnel to label the collected data. In our work, the non-honks data are labeled as ’0’, LWV as ’1”, MWV as ’2’, and HWV as ’3’. Besides audio files, corresponding video files are also used to make our annotation error-free for manual labeling. A total of 1694 discrete samples (1- 3 sec) spanning 30 minutes duration were labeled manually. The details of the manual label data are given in Table 4. We faced some challenges in several cases to recognize the honk of medium-weight vehicles (MWV), as shown in Table 4. For example, the MWV-honking is often labeled as HWV-honking by the personnel. Similarly, the honk of MWV is tagged as LWV-honk in many situations. Results show that our manual labeling process gives a total dissimilarity of around 7%. This dissimilarity might occur due to several issues: 1) raw audio data includes a variety of ambient noises, 2) honk sounds came from different sides of the lane, which is not captured in the videos and the signature of different vehicle honks are sometimes confusing due to their similar pattern, and 3) occasionally, multiple vehicular honks may overlap in time. To improve the accuracy of our training process, we remove mismatched data from our contemplation. However, these remaining samples are too few for training a model, which will result in more errors and low accuracy. Therefore, we require an automated data labeling model to annotate our data with higher accuracy.

Autoencoder(AE) and semi-supervised generative adversarial network (SGAN) based labeled data set preparation has been addressed in many works. The AE-based multi-class data labeling techniques are discussed in Bank \BOthers. (\APACyear2020); Wicker \BOthers. (\APACyear2016); Law \BBA Ghosh (\APACyear2019); Aamir \BOthers. (\APACyear2021). Authors Wicker \BOthers. (\APACyear2016) introduced a multi-class data labeling autoencoder model, where they have compressed the labels using AE by eliminating the non-linear dependencies. A stacked autoencoder network (SAE) is developed in Law \BBA Ghosh (\APACyear2019) to produce a discriminating and decreased input representation of the multi-label data. Similarly, the contractive autoencoder (CAE) model uses layered architecture followed by a feed-forward mechanism to encode unlabeled training data Aamir \BOthers. (\APACyear2021). Apart from these studies, the researchers also developed SGAN-based data labeling models Khan (\APACyear2022); Amin \BOthers. (\APACyear2020). Amin \BOthers. (\APACyear2020) used both SGAN and transfer learning models for labeling the data and model training at a time. In our previous work Maity, Alim\BCBL \BOthers. (\APACyear2022), we considered the SGAN model to label the data into two classes (honk or non-honk) to identify vehicular honks from traffic data in presence of ambient noise. Researchers also tried to combine the autoencoder and adversarial learning for fault diagnosis in Wen \BOthers. (\APACyear2023).

Inspired by this, at first, we tried to label our collected data using the existing autoencoder model, SGAN, SAE Law \BBA Ghosh (\APACyear2019), CAE Aamir \BOthers. (\APACyear2021). Although these techniques give satisfactory outcomes in earlier cited works, they failed to obtain decent accuracy in our work due to the above-mentioned issues. Fig. 7 shows the t-SNE plot of the data samples. To improve the clarity of the plot, we have chosen 30% of the actual data set and noticed that, only in a few cases, the honking signature of LWV and HWV overlapped, but in most of the cases, MWV honking pattern is confusing due to similarities between patterns, either HWV or LWV. Therefore, we need a model that can distinguish all the vehicle types efficiently. In this paper, we have extended the autoencoder (AE) model and proposed two data labeling models, i.e., Multi-label Autoencoder (MAE) and Multi-label Autoencoder Generative Adversarial Network (MAEGAN).

(i) Proposed Multi-label Autoencoder (MAE): Autoencoder is the combination of encoder (squeeze the input into a lower-dimensional code) and decoder (reconstructs the image from the lower-dimensional image) along with the latent space (compressed input). In this work, we have used four different autoencoders (AE0 to AE3), where AE0 is used for the images of labeled 0, AE1 is used for the images of labeled 1, and so on. For all four encoders, four latent space(Z0, Z1, Z2, Z3) is generated. Later we combine all the latent space and create a single latent space Z. Now, we pass this Z to CNN models for the vehicular honk classification. We calculated the probabilities outcome of each class, and the highest probability is considered as a selected class. Moreover, we cross-validate the selected class with the amplitude value of the respective raw audio samples. In this way, we have trained our MAE model for data labeling. Once the training is done, for data labeling, we take the unlabeled data and pass it to the trained encoder, from which latent space is generated for all the classes, and then pass it to the CNN model and select the highest probabilistic class with the cross-validation with corresponding amplitude values. The architecture of the proposed model is illustrated in Fig. 8.

For our proposed MAE, Latent space Z is represented as

where $x_{i}$ is the original input image, $b_{i}$ is the bias, $W_{i}$ is the weight and $\sigma$ is the activation function. Similarly, decoding is represented as

Loss function is formulated as

In all cases, the value of $i$ ranges between 0 to 3.

(ii) Proposed Multi-label Autoencoder Generative Adversarial Network (MAEGAN): A generative adversarial network (GAN) has two components. One is a generator, which generates the fake images, and another one is a discriminator, which distinguishes the real and fake images. In MAEGAN, instead of one generator and one discriminator, we have used four generators (G0-G3) and four discriminators (D0-D3) to label four different classes, i.e., non-honk, LWV, MWV, HWV. In our work, G0 generates fake images for labeled 0, and D0 is used to distinguish the real and fake images of labeled 0. The remaining generators and discriminators are used for other classes of images. The proposed MAEGAN model uses the trained latent space of the MAE model as a generator. The intuition behind using the latent space of the MAE model is that it is already trained with all the different labels of the image classes. Hence, the generator will generate more accurate images, which may be hard to classify by the discriminator. Similar to the proposed MAE model, we have used the same classifier to classify different classes of honks. It is found that the proposed MAE performs better than MAEGAN, which is shown in the result section in detail.

By combining manual and MAE model generated samples, we obtained a total of 56399 samples, where 26%, 20%, and 12% of data belongs to LWV, MWV, and HWV classes, respectively, and the remaining 42% of data is in non-honk class. Due to the disparity in size among honk classes, we faced an imbalanced classification problem in our work. Class balancing in such a scenario may lead to losing a large amount of data, which may further result in the reduction of training samples. To overcome this problem, we do the data augmentation on the spectrogram image rather than raw audio. There exist multiple methods of data augmentation (like time warping, time masking, frequency masking, etc.) to enhance the data volume artificially.

In our work, we have used time and frequency mask techniques to augment the volume of data. In these techniques, some vertical bars are randomly placed to conceal the time masks and horizontal bars to conceal frequency masks on the spectrogram images, respectively. To visualize this, we have considered the spectrogram image shown in Fig. 6. We see the spectrogram image with vertical and horizontal bars showing time mask and frequency mask-based augmentation in Fig. 9 and Fig. 9, respectively. After performing data augmentation, the data volume is increased by around 60%. Next, we do the class balancing, which contains a total of 22 hours of data, and split the dataset for training and testing. The details of the distribution are shown in Table 5.

In several domains, such as object detection, image recognition, and image classification, Convolutional Neural Networks (CNNs) are widely utilized Piczak (\APACyear2015); Salamon \BBA Bello (\APACyear2017). Transfer learning-based CNN models Demir \BOthers. (\APACyear2020) are often employed to achieve higher accuracy. These models leverage pre-training on the ImageNet dataset, which consists of over 14 million images and is accessible through trusted public Keras libraries. By utilizing pre-trained models, training time is reduced, and there is less dependability on having a large training dataset.

(i) Pre-trained CNN models: For classification, we have chosen four pre-trained CNN models, namely MobileNet, ShuffleNet, ResNet 50, and Inception V3. These models were selected by considering two environments: (i) low-end systems and (ii) hardware accelerator systems. We have used two lightweight models for low-end systems: MobileNet and ShuffleNet. On the other hand, ResNet 50 and Inception V3 are used as Heavyweight models for hardware accelerator systems. For our problem, these models have been trained on our dataset. Additionally, pre-trained models are fine-tuned by adding a dense layer to overcome the problem of overfitting.

(ii) Proposed EnTL model: The ensemble model leverages the idea that aggregating the predictions of multiple models can often lead to more accurate and robust predictions than using a single model. To improve the classification accuracy, we propose a transfer learning-based ensemble method, i.e., Ensambled Transfer Learning (EnTL), combining four fine-tuned pre-trained models (i.e., MobileNet, ShuffleNet, ResNet 50, Inception V3). All the models are trained independently on different subsets of the training data, created through random sampling with replacement. Each model provides a probabilistic score for each class using the ’softmax’ function. For an input image $x$, a model can generate four probability scores, one for each class. Let $C_{k}^{i}$ be the probability score of model $k$ on class $i$, where $k$ = 0,1,2, … (M-1); $M$ represents the number of models and $i$ ranges from 0 to 3. We sum up the probability of each class $i$ generated by $M$ models, which is a 1-D vector for $x$. From this vector, we select the class having the highest value using majority voting as a final class for the input image sample $x$ (see in Equation 4). Fig. 10 illustrates the architecture of the proposed model(EnTL).

To evaluate the effectiveness of the proposed data labeling models i.e., MAE and MAEGAN, we perform a comparative analysis against the conventional AE model, GAN, CAE model Aamir \BOthers. (\APACyear2021), and SAE model. All these models are discussed previously in Section 5.2). Table 6 summarizes the experimental results of all models, MAE and MAEGAN. It can be seen that the CAE model performs better in contrast to existing deep learning models i.e., the conventional AE, GAN, and SAE models. But the results indicate that our proposed MAE and MAEGAN models outperform the CAE model. In fact, the performance of MAE and MAEGAN models is $\sim$18% better than the CAE model. This is because the existing algorithms failed to distinguish the honking pattern of MWV and HWV in several cases due to similarities between the data samples, as shown in Fig. 7. As a result, their accuracy dropped. But our multi-label autoencoder models clearly distinguish the honking pattern of MWV, HWV, LWV, and even non-honks and thus give higher accuracy. Furthermore, comparing the performance of the MAE model with the MAEGAN model, we see that the MAE shows $\sim$2.5% better outcome than the MAEGAN. Due to this performance superiority, we have used the MAE model for labeling the unlabeled data in our work. To verify the correctness of the proposed MAE, we have performed a ground truth verification discussed in the following sub-section. A total of 1694 samples are labeled manually, and the remaining data are labeled using MAE. The entire data set is divided into five distinct groups, and we fed each group of data one by one to the MAE model for labeling. A total of 54705 samples are labeled by the proposed MAE model with more than 97 % accuracy. Epoch-wise training and validation accuracy are also depicted in Fig. 11.

To check the correctness of our proposed MAE model, we picked up 340 samples randomly for validation and labeled them using MAE and MAEGAN models. The samples were also fed to other existing data models (AE, GAN, SAE, CAE) for labeling and then compared with MAE and MAEGAN models. For ground truth verification, these 340 samples are further labeled manually by two volunteers, where the number of LWV, MWV, and HWV honking samples are 88, 64, and 39, respectively. The results are shown in Fig. 12. We see that only one honk of MWV is misclassified as the honk of HWV, and 8 HWV samples are detected as LWV honks by the MAE model. It implies that out of 340 samples, only 9 samples are wrongly tagged by the proposed MAE model, while the misclassified samples generated by other models are 16, 115, 150, 139, and 166 for MAEGAN, AE, SAE, CAE, SGAN models respectively. The result signifies that our proposed MAE model can correctly label the unlabeled samples.

Table 7 illustrates the performance comparison for all the pre-trained models and EnTL for classifying honks using the chosen performance measures. Here, we have used 79608 labeled samples, where 63684 samples were for training and 15924 samples were for testing. We have trained all pre-trained models using the Adam optimizer for 20 epochs. The input image size is 224 × 224 × 3, the learning rate is set at 0.001, and the batch size is set to 312. Results indicate that lightweight models i.e., MobileNet and ShuffleNet, have achieved test accuracy at around 88.82% and 78.59%, respectively.

On the other hand, ResNet 50 and Inception V3 have achieved accuracies of 93.77% and 95.17% , respectively. among the pre-trained models, Inception V3 gives better results in our dataset. Nevertheless, the proposed EnTL model surpasses that of Inception V3 with an accuracy of 96.72% and demonstrates classification accuracy of 95.47%, 97.32%, and 97.72% for LWV, MWV, and HWV, respectively. Additionally, the attained values of MCC, F1-score, precision, recall, Kappa statistics, and ROC-AUC demonstrate that EnTL outperforms other models. The confusion matrix, shown in Fig. 13, is calculated for each model. We have found that the misclassification (false positive) rate is significantly less in EnTL. Only a few non-honk samples (2.7%) are detected as LWV because sometimes the pattern of engine noise/several other noises becomes similar to the LWV. Similarly, 2.41% samples of LWV are classified as MWV due to the same type of honking pattern that may arise in the case of motorbikes and four-wheeler vehicles. The remaining misclassification rate is less than 1%, which is negligible. As a note, the classification performance of the proposed EnTL is better than any single transfer learning model and can be used as a model to classify vehicular honks even in the presence of ambient noises.

To show the importance of our curated dataset, we have trained the best-performing model EnTL for different datasets and plotted the training accuracy in Fig. 14. At first, we trained the model with manually labeled samples. The training accuracy dropped since our manually labeled samples(1694) were few. Next, we trained the model using MAE-generated labeled samples (54705) and obtained better accuracy than training with manual label samples. Finally, the model was trained with the enriched dataset(79806) generated by data augmentation techniques, and we achieved a very good training accuracy. Hence, our proposed methodology and the dataset play a crucial role in this study as well as the annotate dataset can be useful for other research.

In this section, the performance of the EnTL model is compared with baseline work using chosen performance metrics. Here, we have considered SBCNN Salamon \BBA Bello (\APACyear2017), Dilated CNN Chen \BOthers. (\APACyear2019), CNN Demir \BOthers. (\APACyear2020), and TFCNN Mu \BOthers. (\APACyear2021) as baseline works. The configuration settings of each baseline model are represented in Table 8. The detailed results are presented in Table 9, from where we notice that the performance of EnTL is increased by at least $\sim$9% and at most $\sim$21% in contrast to other models. The baseline models reported higher sound classification accuracy in publicly available datasets like ESC-10, ESC-50, or UrbanSound8k, where ambient noise was either absent or not considered. Moreover, each audio file’s duration is much less in these datasets. In contrast, our data was collected from a real environment in the presence of various ambient noises. Hence, the accuracy fell sharply when we fed our dataset to the existing models. However, the Dilated CNN provides the best output among all the baseline models. When we compare our model with Dilated CNN model, the performance of our model is increased by 10.15%, 17.07%, 11.49%, 10.22%, 12.79%, 17.07%, 7.69% for accuracy, MCC, F1-score, Precision, Recall, Kappa Statistics, ROC AUC. The EnTL model is good for classifying honk based on the vehicle types in the presence of ambient noises.

• •

  Importance of our curated dataset: Due to the fundamental nature of the traffic data, our dataset was the combination of several ambient noises, but still, we achieved a decent accuracy for honk classification. When we performed the same experiment with the baseline models, the overall accuracy decreased as all the baseline models were trained either in a controlled environment or ambient noise-free data only. Moreover, we have generated 22 hours of labeled data repository, which can further be used for developing several micro-services based on honk features.

• •

  Significance of our proposed data labeling model: Due to the similar pattern of honk signature, the existing data labeling models failed to reach a good accuracy, whereas our proposed MAE model succeeded with 21% more accuracy than existing models. Therefore, we can extrapolate that this proposed model is more suitable and will give better outcomes when the data pattern is very closely related.

• •

  Role of honk classification models: There may be circumstances where real-time detection and model execution are both required on some resource-constrained devices like IoT edge devices or smartphones. In this situation, we suggest using the MobileNet model because it is lightweight and doesn’t significantly degrade performance (by about 6.53% as seen in our tests). Furthermore, our proposed model EnTL outperforms other models and provides the best outcome.

• •

  Inference of context: Proper honk classification facilitates us in recognizing the outdoor context of a location. It has had a big influence on researching passive sensing methods to develop several micro-services. In this research, we have presented a glimpse of spatio-temporal context of three different locations.

To assess the context of a location, we have collected vehicular honk data from different outdoor scenes i.e. highways, market place, and residential areas for three different time slots. Continuous ten days of data have been collected for the five minutes duration of each time slot. The overall hoking signature for all the locations and time slots are depicted in Fig. 15. We have also provided a snap of all the places that were visited in different time slots. From Fig. 15, we have observed the following sound pattern (f# denotes the features number for each of the locations):

• •

  Highway: [f1] Throughout the day, the highest number of honks are generated by HWV. [f2] The number of honks generated by the LWV is the lowest as compared with others. [f3] All vehicle honks increase as time proceeds.

• •

  Market place: [f1] Honks of LWV and MWV are higher. [f3] comparatively, HWV honks are much lower. [f2] Overall honking pattern goes down in the afternoon.

• •

  Residential Area: [f1] Some LWV honks are found throughout the day [f2] HWV honks are almost negligible in this place. [f3] In comparison to other places, the number of honks is significantly less in residential areas.

Therefore, we can say that a unique temporal pattern exists in all the locations, which distinguishes one place from another.

In order to verify how well our proposed system $\mathcal{A}$$\mathcal{C}$lassi$\mathcal{H}$onk works, we moved through four different areas in a single trace and plotted the various honk data in a graph. The perceived result is presented in Fig. 16 and the street view of our travel is also shown along with the latitude and longitude value in Fig. 17. We started our journey from the residential area and then passed through a marketplace, low-traffic area, high-traffic area, and again returned back to the residential area. Fig. 16 shows all most similar kinds of honking patterns that we observed in Section 7.1. A distinguishable honking pattern among all the locations is visible, which justifies our institution to identify the context of a place.