Damage Detection in the Blink of an Eye

Damage Detection in the Blink of an Eye
A Leading Car Rental Company Leveraged Computer Vision for Incremental Damage Detection in less than 100 Seconds

Common Page CSS / JS

About the Client

The client is a leading American car rental agency holding company headquartered in Parsippany, New Jersey. It operates across 180+ countries globally with a revenue of approximately $12 billion USD. They needed to improve overall time to identify damage across vehicles and optimize the overall customer experience.

Need for Change

In the car rental industry, there are several challenges driving the need to harness innovative technology such as AI and computer vision to improve the speed and accuracy of the damage detection process and optimize the overall customer experience. False positives, low accuracy and difficulties in distinguishing new damage from pre-existing damage are some of the factors fueling this transformation. AI integration in car rental visual inspection processes will streamline and automate damage detection processes with greater accuracy and efficiency. In the near future, AI car inspection will be the norm as more companies adopt this technology.

Business Challenges

No process to distinguish new damages from pre-existing ones, leading to inaccuracies in damage identification. Addressing incremental damage detection was a key challenge.
Lack of automation and streamlined technology for damage identification resulted in significant manual efforts and high upfront costs.
Inability to capture and accurately classify major and minor damages based on the vehicle area.
The existing solution had low accuracy and a high false positive rate.

Our Approach

We started by aggregating data at the source to compare the check-in and check-out videos for incremental damage identification.

Once the data was captured and available, by leveraging the Vision NXT CV platform, we developed multiple AI-driven models to detect and identify damages and identify the precise vehicle part where damage had occurred.

The end results were shared in less than 180 seconds to the customer’s source system, a damage management portal (DMP) for reporting multiple downstream and upstream services.

Camera + cloud + data processing + AI model + results (JSON) 🡪 customer’s source system (DMP)

Tech stack

AWS

Data storage layer for RAW video feeds for model training purposes

OpenCV

Image processing library for video analysis and pre-processing video feeds data

Python

Underlying backend stack for API development, custom business logic, and outcomes generation

Elasticsearch

Embedding store and similarity search for reidentification of vehicle damages

Deep Learning

Deep learning algorithm as the underlying model training mechanism to identify vehicles, damage detection, and segmentation of the vehicle parts

Triton Server

Orchestration and inferencing layer to infer vision model outcomes based on a given timeframe

Solution Highlights

We collaborated with the client as their solution partner using an AI-driven approach leveraging our Vision NXT CV platform for their transformation journey. This AI integration in the car rental company’s visual inspection processes included the following key steps:

Discovery phase

We identified the key pain points, the hardware (camera) set up for real-time video feeds, and their existing downstream services (to seamlessly integrate the end results) for their team.

Application design

We created a modular and microservices-based architecture tailored to the following:

Video aggregation: Aggregated 10 videos for the same vehicle as a transaction for processing
Vehicle detection: Analyzed video feeds in real-time to detect vehicle presence
Parts segmentation: Segmentation model to identify the correct exterior part of the vehicle
Damage identification: Damage detection model to flag, detect, and identify the type of damage
Incremental damage: ReIdentification (ReID) model to differentiate between new and pre-existing damages
Back-end processing and data structuring: Structured the end outcomes as JSON to integrate with the customer’s downstream services and systems

AI-enhanced computer vision models

Developed and deployed vision-based models for AI car inspection to improve overall incremental damage identification speed to < 180 seconds.

Real-time updates

We provided real-time updates on the video feeds (based on the unique vehicle) to the customer’s downstream service (DMP system) in less than 180 seconds.

Advanced analytics reporting

Advanced analytics reporting with exhaustive KPIs and metrics highlighting the overall volume of data processed, incremental damages identified, and much more.

Benefits

Two million+ videos processed with minimal latency

The scale of the solution and architecture was capable of processing more than two million feeds across multiple sites seamlessly with minimal latency and information loss.

More than 90% accuracy achieved to detect and identify damages

Achieved a model accuracy with business acceptance of greater than 90% across multiple models.

More than 80% consistency to track the damages

The models ran with a consistency of more than 80% ensuring the right detection and classification of damages is achieved.

Less than 100 seconds average time

Achieved an average time of less than 100 seconds to generate and send the end results to the customer’s damage management portal.

Four stations

Deployed and the model across multiple sites ensuring variations in multiple conditions such as lighting, weather, traffic, and more.

Conclusion

The AI-powered digital transformation delivered by LTIMindtree empowered the client to detect, identify, and classify incremental damage using computer vision to enhance both user experience and operational efficiency. This enabled faster response times, higher user adoption, and improved process and operational productivity. We will see widespread adoption of AI car inspection and computer vision to automate and optimize such processes across the car rental and automative industries.

Applications across other industries and potential use cases:

You can learn more about LTIMindtree’s iNXT here.

Contact us