Vehicle Interior Detection (2)

▤ 목차

Proof of Concept

728x90

반응형

Research and Development

Picture from https://www.socar.kr/

This section presents the configuration of the project, detailing the data collection strategies, technical methodologies, anticipated outcomes, and a structured implementation plan. The research and development phase is integral to transforming the theoretical concept into a practical, AI-driven solution that addresses the identified challenges effectively.

Data Collection and Categorization

Effective AI systems rely on extensive, high-quality datasets to achieve accuracy and functionality. For this project, two primary categories of data were identified: interior vehicle images and vehicle dashboard images. Each category is purposefully designed to address a unique aspect of the problem—ensuring the cleanliness and mechanical functionality of the vehicles.

 

Interior Vehicle Images

The first category focuses on evaluating the interior cleanliness of vehicles. By collecting images of specific areas prone to dirt and clutter, the AI model will learn to detect issues with vehicle hygiene. This dataset is further divided into two subcategories:

1. Clean Version:
   • Images in this category depict various parts of the vehicle interior in a pristine state. For example:
     • Front left car seat area: This represents the seating area frequently occupied by drivers.
     • Rear right car seat area: This ensures that the back seating is accounted for in cleanliness checks.
   • These images set a baseline standard for what is considered "clean" and serve as the reference for training the AI model.
2. Dirty Version:
   • This subcategory contains images showing dirt, clutter, or damage in the vehicle interior. Common areas include:
     • Armrest: Often used and prone to visible stains and trash.
     • Foot mat area: Frequently collects debris and dirt from shoes.
     • Left backseat area: Represents the rear of the vehicle, where less oversight may occur during cleaning.
   • The inclusion of varied examples in this subcategory ensures the AI model can recognize dirt under different lighting conditions, materials, and severities.

Vehicle Dashboard Images

The second category focuses on identifying potential mechanical issues by analyzing vehicle dashboards. Like the interior images, this dataset is also divided into two subcategories:

1. Without Faults:
   • Images in this subcategory show dashboards free of warning signs, such as those captured immediately after vehicle maintenance. These serve as a reference for normal vehicle conditions.
2. With Faults:
   • This subcategory contains images where dashboard indicators signal potential problems. Examples include:
     • Engine Malfunction: Highlighting issues that require immediate attention.
     • Low Tire Pressure: A common issue that could compromise safety if ignored.
     • Oil Pressure Warning: Indicative of critical maintenance needs.
   • By training the AI model on this dataset, the system can automatically identify faults and notify operators for timely action, ensuring vehicles are safe and operational for the next customer.

Data Collection Methods

To build these datasets ideally, a combination of user-generated content and company-driven efforts will be employed:

• User Submissions: Clients will upload photos through the Socar app during the vehicle return process. These images will serve as real-world data for training and refining the AI model.
• Internal Inspections: Socar’s personnel will supplement the dataset with additional images collected during routine maintenance and inspections. This ensures data diversity and improves model robustness.

If both options are not feasible, data crawling will be performed to obtain license-free images accordingly.

 

In addition, the collected images will undergo preprocessing, including resizing, normalization, and augmentation, to standardize quality and enhance the model’s learning capabilities.

Technical Approach

To develop a reliable AI model, selecting an appropriate methodology is critical. The team initially explored advanced techniques such as semantic segmentation and object detection. While these methods offer greater precision, their implementation poses challenges due to the time-intensive process of data annotation and model complexity.

 

Instead, the team opted for a classification model, which categorizes images into predefined classes (e.g., clean/dirty or fault/no-fault) and an open-set recognition model that is similar to the classification model but categorizes undefined data as unknown. This simpler approach aligns with project timelines and resource availability while still achieving the desired functionality.

Model Workflow

The model follows a structured workflow:

1. Data Preparation:
   • Collected images are labeled according to their respective categories and subcategories.
   • The dataset is split into training, validation (if enough data is collected), and testing sets to ensure unbiased evaluation.
2. Training the Model:
   • The model is exposed to labeled data, allowing it to learn features that distinguish each class. For example, it learns to identify trash patterns in interior images or recognize specific dashboard warnings.
3. Validation and Iteration:
   • The model’s performance is tested using unseen validation data. Metrics such as accuracy, precision, and recall are evaluated, and the model is fine-tuned to address any deficiencies.
4. Deployment Readiness:
   • Once trained and validated, the model is prepared for integration into the Socar app. Continuous monitoring will ensure the model adapts to new data and maintains performance over time.

This streamlined approach balances functionality with practicality, allowing the project to progress efficiently without sacrificing quality.

Expected Outcomes

The integration of AI into the Socar app is expected to deliver significant benefits across multiple dimensions.

 

Short-Term Benefits

In the short term, the implementation will:

• Enhance User Experience: By ensuring vehicles are clean and functional before each rental, customers will have a positive first impression of the service.
• Reduce Complaints: Automating the detection of cleanliness and faults minimizes the likelihood of customer dissatisfaction.
• Streamline Operations: Automating inspections reduces the reliance on manual checks, freeing up personnel for other tasks.

To validate these benefits, A/B testing will be conducted, comparing user feedback and operational metrics before and after the implementation of the AI model.

 

Long-Term Benefits

Over time, the AI system will contribute to:

• Increased Customer Loyalty: Consistently clean and functional vehicles will foster trust and encourage repeat usage.
• Operational Efficiency: The reduction in manual inspections and improved resource allocation will lower operational costs.
• Brand Differentiation: By leveraging AI to enhance service quality, Socar positions itself as an innovative leader in the car-sharing market.

The long-term impact extends beyond customer satisfaction, strengthening Socar’s competitive advantage and profitability.

Action Plan

To ensure a systematic implementation of the project, a detailed action plan has been developed. The plan is divided into three phases:

1. Data Collection and Preprocessing:
   • Collaborate with users to gather diverse datasets.
   • Conduct internal inspections to supplement user submissions.
   • Preprocess images to standardize quality and improve model training efficiency.
2. Model Training and Evaluation:
   • Train the AI model using labeled datasets.
   • Validate performance and refine the model based on feedback and results.
3. Optimization and Deployment:
   • Integrate the model into the Socar app and test it in a controlled environment.
   • Collect user feedback during beta testing to identify areas for improvement.
   • Deploy the model at scale, with continuous monitoring and iterative updates.

This phased approach ensures the project is completed on schedule and delivers the desired outcomes effectively.