Micromobility annotation

Micromobility annotation is the labeling of cyclists, pedestrians, scooters, skateboards, bicycles, and other micro-mobility devices for training urban mobility AI in complex urban scenarios, including accurate e-scooter detection.

The main goal is to transform complex street scenes into structured data through bounding boxes, keypoints, segmentation, and tracking, enabling reliable e-scooter detection and understanding of micro-mobility devices in real-world urban scenarios.

Quick Take

• Annotate occupants and devices using accurate, reproducible standards.
• Contextual tags, lanes, crossings, and lighting improve safety modeling.
• Hybrid workflows combine automated annotation with expert review.
• Outputs are aligned with downstream uses: detection, prediction, and policy.

Abstract of micromobility

Annotation of electric scooters, mechanical scooters, skateboards, and pedestrians is an important component for the development of modern computer vision systems, smart cities, and autonomous transport, supporting e-scooter detection and behavior modeling of micro-mobility devices in dense urban scenarios. Micromobility requires a separate taxonomy that allows models to distinguish between road users and their behavior correctly.

The taxonomy of micromobility is built hierarchically. First, the class of "vulnerable road users" is distinguished, including pedestrians and users of personal mobile means of transport. Then the objects are detailed by type. Within each class, additional separation is possible by pose and state: moving, standing, rolling next to each other, riding on the roadway, on the cycle path, or on the sidewalk.

Areas where micromobility annotation is used

In autonomous transportation and ADAS systems, such data is needed to identify vulnerable road users, enable robust e-scooter detection, predict trajectories of micro-mobility devices, and reduce collision risk in complex urban scenarios.

In the field of smart cities, it is used to analyze traffic flows, plan infrastructure, and optimize the placement of bike lanes and shared traffic areas.

For scooter and skateboard rental services, annotated data helps track usage patterns, identify dangerous scenarios, and improve operating rules.

In video analytics and security systems, micromobility is essential for accurately interpreting human behavior in public spaces, where the distinction between pedestrians and vehicles is often blurred.

The data needs for micromobility annotation are high due to the variety of scenarios and the complexity of visual recognition. Datasets with different camera and sensor types, viewing angles, lighting conditions, weather, and traffic density are needed.

It is vital to collect data from city centers, residential areas, and transport hubs, where micromobility user behavior varies significantly. Annotations should not only be spatial (bounding box, segmentation), but also semantic and behavioral.

Safety, equity, and policy cues that your dataset should capture

The dataset should capture not only the physical objects and their movement, but also the context of safety, equity, and policy cues that influence the models' interpretation of the scene and decision-making. Such cues form the "social layer" of the data and are essential for creating technologies that work correctly in the real world.

Safety cues

Safety cues include all visual and behavioral indicators that indicate potential risks or safety conditions. These include road signs, traffic lights, markings, crosswalks, speed limits, and temporary features (such as cones, fences, construction barriers, and emergency signs). Personal safety markers are also captured, including helmets, reflective elements, protective clothing, and pedestrian or traffic signal gestures.

Special attention should be paid to dangerous scenarios: sharp maneuvers, rule violations, close passages between traffic participants, sudden stops, or trajectory intersections. These are the situations that are important for training models to prevent such incidents.

Equality signals

Equality signals ensure the representativeness and fairness of the data and capture the diversity of contexts and conditions in which people are.

The dataset should reflect different body types, age groups, clothing styles, and the use of mobility aids (canes, walkers, wheelchairs, baby carriages).

It is also important to consider different accessibility scenarios in infrastructure, such as ramps, lowered curbs, tactile tiles, sidewalk widths, and obstacles in the path of movement. This approach allows models to work reliably for all users of urban space and reduces the risk of systemic bias.

Policy and regulatory signals

Policy and regulatory signals establish rules and restrictions at the city or state level and set the framework for acceptable behavior in a given space.

These include zones where micromobility is allowed or prohibited, pedestrian zones, shared spaces, bike lanes, and temporary restrictions during mass events or emergencies.

It is also important to capture information signs, signs indicating local regulations, and markings for scooter parking and docking areas. In the data, these signals are associated with the behavior of traffic participants, so that models can not only see the rules, but also understand when they are violated.

Taken together, all this forms a contextually rich dataset that allows systems to train not only to recognize objects but also to interpret the environment as a socially and normatively structured space.

How accurate labels reduce risk

Insurance and risk management depend on the quality of the data used to train and operate autonomous systems.

Accurate annotation allows for the reconstruction of events and the interpretation of participant behavior in complex scenarios. When a system has annotated data about the types of objects, their states, movement trajectories, and regulatory context, it is easier to determine whether the algorithm's decision was justified.

Insurance models increasingly rely on computer vision and sensor data to estimate the likelihood of incidents and set insurance rates. Annotation enables distinguishing between high- and low-risk scenarios, classifying participant types and environmental conditions, and analyzing the frequency of dangerous events. This reduces the likelihood of erroneous payments or disputes between the insurer and the system operator, as decisions are based on transparent, reproducible data rather than assumptions.

For management, accurate labels help move from a reactive to a proactive approach. When data clearly shows dangerous patterns, regulatory violations, or weak infrastructure, management decisions are made on the basis of facts, not intuition. This applies to urban planning, micromobility fleet management, or autonomous systems.

Choosing a data annotation provider

To achieve high-quality AI models, you need to choose the right annotated data provider. Keymakr specializes in image and video annotation, data creation, validation, and classification for AI/ML systems. A large team of specialists supports four levels of quality assurance and adapts solutions to the specific needs of each project. Its proprietary Keylabs platform provides a user-friendly interface, supports various annotation types, and project management tools.

Keymakr has worked on multiple projects that required micromobility annotation. In a public case study conducted under NDA conditions, Keymakr annotated road scenes and videos for training driver-assistance systems (ADAS) and road-object recognition models. The project required high accuracy, processing of large amounts of data, and adaptation to different regional traffic conditions. During the first phase of work, the team processed up to 200,000 object instances in a short time. In subsequent stages, the requirements for scalability and quality increased significantly, leading to the decision to migrate the work to the Keylabs platform. This enabled more than 50 annotators to work simultaneously without loss of productivity, resulting in an eightfold increase in throughput. In total, more than 500,000 data units were processed, which significantly supported the development of systems with a high level of real-world application.

Problems and solutions in micromobility annotation

Accurate annotation allows algorithms to correctly recognize objects, predict movement trajectories, and ensure road safety. However, specific problems arise from the diversity of objects, user behavior, and urban environmental conditions. These challenges are especially critical for e-scooter detection and other micro-mobility devices operating in highly dynamic urban scenarios. Let's consider how to solve them:

FAQ

What is the scope of micromobility annotations for e-scooter detection and micro-mobility devices in urban mobility datasets?

The scope is to improve the safety and accuracy of autonomous transportation systems, analyze urban flows, and plan infrastructure.

What label sets are important for safety and policy analysis?

Labels of road signs, markings, traffic lights, traffic zones, pedestrians, micromobility users and their behavior. They reflect rules and potential risks.

How do annotated datasets improve accident reconstruction and liability assessment?

Annotated datasets enable accurate reconstruction of object trajectories and event circumstances, thereby increasing the reliability of accident reconstruction and the objectivity of liability assessment.

What attributes should be captured to support safety interventions?

The attributes of objects, their position, speed, direction of movement, presence of protective equipment, and behavioral signals that can influence risky situations should be recorded.

What are the main problems in micromobile annotation?

The main problems are related to the diversity of objects, fast and unpredictable user behavior, mixed space use, and difficult visibility conditions.