1 Introduction

Two-wheelers are a common form of transportation in India but are also frequently involved in serious accidents and crimes. As a result, research on smart traffic systems, including vehicle detection, identification, enumeration, and traffic statistic estimate, has become increasingly popular. Researching traffic estimates, speed calculations, motorcycle helmet use, vehicle tracking, as well as occlusion analysis requires segmenting automobiles on public roadways.

The proposed work aims to address the challenge of detecting and classifying motorbikes on public roads. As our world moves faster, the number of vehicles on the road is increasing daily, along with a corresponding rise in crimes and accidents involving bikes. This increase appears to be linear.

To find who is the cause of the accident or the culprit in a crime. The investigation officer enquires about the people in and around the place where the incident has taken place. but they won’t have recognized or seen the person’s face or the number plate they may have seen only what color the bike was and how did the bike look like or the color of the bike.

Hence to find a solution to this problem, the proposed work is to find the bike model using the bike’s wheel rim and bike headlamp. we could have used the number plate to identify the bike and trace out the person who rode the bike but sometimes the number plate may be damaged or sometimes the numberplate may be slightly faded and which is not visible at all or sometimes the bike maybe without number plate also hence without wheel rim a bike cannot move or is not complete hence with the help of wheel rim and headlamp we are going to detect the bike model. Figure 1 shows a sample image of the bike with localization of the wheel rim and headlamp.

Fig. 1 Sample images from the collected data set 

2 Related Works

Research into computer vision and image processing (IP) techniques for automating vehicle classification has been ongoing.

One possible approach involves pre-processing images using IP techniques to enhance their suitability for input into a machine-learning model. For example, in [¹], a stationary camera was used to record footage of vehicles, which were then classified using three levels of processing and appropriate cameran calibration. Similarly, [²] suggested using IP techniques, Invariant Feature Transform (IFT), K-means clustering, and euclidean distance matching to classify vehicles into different categories for Electronic Tolling Collection.

Another example is a pre-crash detection system, as described in [³], which utilized a front-facing low-light camera to detect vehicles and avoid accidents. This system used a combination of multiscale-driven hypothesis generation and appearance-based hypothesis verification to achieve good results in all kinds of weather and road conditions.

In [⁴], Sonoda et al. proposed a system for detecting moving objects, such as vehicles and pedestrians, at an intersection, which warns drivers of potential dangers. The system employed a mixture of Gaussians for object detection and the Lucas-Kanade Tracker (LKT) algorithm for pedestrian tracking.

In [⁵], researchers used a camera to capture images of vehicles against a static background, allowing them to calculate vehicle length and width and classify the vehicle as either a Light Motor Vehicle (LMV) or a Heavy Motor Vehicle (HMV). The goal of this research was to determine a suitable toll rate.

In [⁶], an automated vehicle classification system was proposed to distinguish between motorcycles and sedans, using commercially available light curtains to produce images and two machine learning methods, K Nearest Neighbor (KNN) and decision tree learning, to classify the vehicles. Finally, [⁷] proposed a computer vision system to detect and segment motorcycles that were partially occluded by other vehicles, using a helmet detection system to identify motorcycles.

However, this approach requires significant information, such as helmet radius, camera angle, and camera height, to function effectively [⁸] developed a system that can identify vehicles even if they are partially occluded and can continue tracking the vehicle despite occlusions. In [⁹], a computer vision system was developed for detecting bicycles, pedestrians, and motorcycles.

This system used Gabor filtering to detect motion and a Histogram of Oriented Gradient (HOG) descriptors to classify two-wheeled vehicles and pedestrians. Chiverton et al. [¹⁰] proposed a system for automatically classifying and tracking motorcycle riders with and without helmets, using histograms derived from the head region of riders to train an SVM classifier.

Although this system achieved high accuracy, the number of test images was limited. The current research work mainly focused on vehicle classification with limited classes and quantity. But there is a need for the bike model dataset with large classes for automation of the bike model identification.

This proposed work created a bike model dataset and pre-trained the Faster Region Convolution Neural Network (RCNN) model for the identification of the bike model. The major contributions of the proposed method are:

3 Proposed Methodology

The proposed methodology comprises four stages: database creation, augmentation, and training and testing phases to identify motorcycle models using categorical images. The dataset was created by collecting images from various sources LIKE search engines, movies, CCTV footage, and bike showrooms, resulting in a diverse collection of two wheelers.

Data cleaning was performed to remove unwanted images, resulting in a dataset of 1,500 motorcycle images from 27 categories, which were augmented to obtain a total of 5,000 distinct images stored in the database.

However, the target variable showed an imbalance in the number of instances across different classes. Sample images from the created bike dataset are presented in Figure 2. The Faster RCNN is a widely used object detection framework that was introduced by Ross Girshick in 2015.

Fig. 2 Examples of annotated bike categories are presented, demonstrating variations in both form and pose 

This architecture employs Convolutional Neural Networks (CNNs) for object detection, similar to other popular designs such as You Look Only Once (YOLO) and Single Shot Detector (SSD).

When given an image and bounding boxes as input, the Faster R-CNN network processes the entire image using multiple convolutional and max pooling layers to generate a convolutional feature map.

In this study, the Faster RCNN was used to detect and identify objects in the created bike dataset. To generate feature maps of sizes 60, 40, and 512, the image is passed through the ResNet backbone of the CNN.

For proposal generation, the Region Proposal Network (RPN) is utilized, which benefits from weight sharing between the Faster R-CNN detector backbone and the RPN backbone.

The input image is resized to 640×640 and is fed into the CNN’s backbone to initiate the RPN. This study introduces backbone networks like ResNet, which have a network stride of 16, leading to significantly smaller output features compared to the input image. Each point on the output feature map corresponds to two points 16 pixels apart in the input image.

The proposed deep detector network aims to identify whether there is an object present in the input image at each point in the output feature map and estimate its size.

To achieve this, a set of "Anchors" is placed on the input image for each position on the output feature map generated by the backbone network. These anchors point to potential objects of various sizes and aspect ratios that could be present.

The network then determines which of the k-associated anchors in the input image contain objects and adjusts their coordinates to provide bounding boxes as "Object proposals" or areas of interest.

To create 512-d feature maps for each location, the backbone feature map undergoes a 3×3 convolution with 512 units.

Two sibling layers are added: a 1×1 convolution layer for object classification with 18 units and a 1×1 convolution layer for bounding box regression with 36 units. The classification branch's 18 units output a probability of whether an object is present at each location in the backbone feature map.

The regression branch's 36 units produce the size of the object proposals for each point in the backbone feature map. The regression coefficients of each of the 9 anchors for each point in the backbone feature map are obtained using this output.

The coordinates of the anchors containing objects are then adjusted using these regression coefficients

The backbone feature map features are obtained using the ROI pooling layer, which pools them based on RPN-proposed bounding box boundaries. The ROI pooling layer selects the area that matches the proposal from the backbone feature map, divides it into a fixed number of sub-windows, and performs max-pooling over these sub-windows to produce an output of a fixed size.

The size of the ROI pooling layer's output is (N, 7, 7, 512), where N is the number of proposals generated by the area proposal algorithm.

These features are passed through two fully connected layers before being fed into the sibling classification and regression branches. Two objects need to be labeled in a bike model detection task, and the labels for each object can be either 'headlamp' or 'wheel rim'.

The weights assigned to each label are W1 = 0.6 for 'headlamp' and W2 = 0.4 for 'wheel rim'. We assign a higher weight to the headlamp (0.6) than to the wheel rim (0.4) as the headlamp is more significant than the wheel rim. There are two possible label conditions:

To determine the label for an object, we use the following process:

Then compare the values of W1S1 and W2S2 to determine which label is more appropriate for the object. If W₁S₁ is greater than W₂S₂, then the object is labeled as the label described by the headlamp, and if W₂S₂ is greater than W₁ S₁, then the object is labeled as the label described by the wheel rim.W1S1 h1h0≷W2S2.

4 Results and Discussions

The results and discussion of the proposed approach are described in this section. The collection of bike images that were developed was divided into an 80:20 ratio. 80 percent of the samples were chosen for training, 10 percent for validation, and the rest for testing.

The proposed algorithm was trained using Google Colab, which was set up on an HP server with 8GB of RAM. The training images were reduced in size at the pre-processing stage to 640×640. The Faster RCNN pre-trained model was then trained using the resized image. Table 2 contains the default settings for the training parameters of the proposed model.

To increase accuracy, these parameters are modified, including the optimizer, learning rate, and epochs. Figure 3 depicts the proposed pre-trained model's training plot. This model's training recognition accuracy improves to 94.35% once these parameters are adjusted. Table 3 represents the performance metrics of a FASTER RCNN that is designed to detect and classify bikes using the wheel rim and headlamp of a motorcycle.

Average Precision (AP), measures the accuracy of the model in terms of its ability to correctly identify the presence of a bike, as well as its ability to correctly identify the specific type or model of the bike.

The value of 0.9435 indicates a relatively high level of accuracy. Average Recall (AR), indicates the proportion of actual bike instances that were correctly identified by the model.

The value of 0.738 suggests that while the model is accurate, there is room for improvement in terms of correctly identifying all instances of bikes in the dataset. Localization loss measures the error in the model's ability to accurately localize the object in the image. The value of 0.065122 indicates that the model is performing well in this regard.

The classification loss measures the error in the model's ability to accurately classify the type or model of the bike. The value of 0.167798 suggests that the model is performing reasonably well in this regard, but there is room for improvement. Regularization loss measures the degree of regularization in the model, which is designed to prevent overfitting.

The value of 0.000000 suggests that the model does not overfit the training data. Total_loss is the sum of the three previous loss values and represents the overall error of the model.

The value of 0.275017 indicates that the model is performing well, but there is still room for improvement in terms of correctly identifying all instances of bikes in the dataset, as well as accurately classifying the type or model of the bike.

Figure 5 shows a decreasing trend in the loss as the number of steps increases, with some fluctuations due to noise in the data or variations in the optimization process.

Fig. 4 Flow of faster RCNN architecture 

Fig. 5 Plot of localization loss vs step 

At the beginning of training, the localization loss is high and the model's predictions are poor, as indicated by the high point at the left side of the graph. However, as the training progresses, the loss decreases steadily, with occasional fluctuations, until it reaches a minimum.

Figure 6 shows the total loss decreases over the course of training, as the model becomes better at making predictions.

Fig. 6 Plot of total loss vs step 

The graph shows some fluctuations in the loss due to noise in the data or variations in the optimization process.

At the beginning of training, the loss is high because the model's predictions are poor.

However, as the training progresses, the loss decreases steadily, with occasional fluctuations, until it reaches a minimum.

In Figure 7, the learning rate schedule for the training process is illustrated. Initially, a high learning rate, such as 0.04, is set to allow the optimization algorithm to quickly search the parameter space and find a good initial solution.

Fig. 7 Plot of learning rate vs step 

During the course of training, the learning rate gradually decreased based on a pre-defined schedule. In the given example, the learning rate is reduced by a factor of 10 after every 1000 steps.

Therefore, after 2000 steps, the learning rate is reduced to 0.03, and after 3000 steps, it is further reduced to 0.02, and so on. The proposed work's performance can be evaluated using the confusion matrix, as shown in figure 8 And the labels of each model are shown in table 4.

Fig. 8 Confusion matrix of bike model classification 

Precision, recall, and the F1 score are used to evaluate the proposed model performance. There are 27 classes, each class has two labels hence 54 subclasses in these 54 subclasses in 54 subclasses 48 subclasses have accuracy above 90%. The proposed model outperforms the others since 94.35% of the Bike models were correctly classified in the testing phase.

Table 5 shows the accuracy of each model on the task, as a percentage. MobileDET achieved an accuracy of 89.5%, YOLO v5 achieved an accuracy of 90%, and the proposed network using Faster R-CNN achieved the highest accuracy of94.35%.

5 Conclusion

Detecting and classifying motorbikes on public roads is a challenging task, but it has become increasingly important due to the rising number of bike-related crimes and accidents. The proposed work focuses on identifying the bike model using the bike's wheel rim and headlamp, as these components are often visible even if the number plate is damaged or not present.

The proposed work consists of bike model database creation, training the bike model by Fater RCNN, and testing the bike model identification.

This work achieved better identification accuracy of 94.35%. The proposed Fater RCNN model compared with YOLO v5 and MobileDET. Ultimately, this work has the potential to make our roads safer and help law enforcement agencies in their efforts to prevent and solve crimes.