Ultrasonic Data Annotation for Parking AI

Currently, automotive autonomous and semi-autonomous driving systems rely on a complex suite of sensors that includes optical cameras, radars, and high-precision LiDARs. Despite their significant operating range and ability to build detailed three-dimensional maps of space, these technologies prove vulnerable in the immediate vicinity of the vehicle's bumper due to the presence of structural blind spots and the high cost of short-range optics.

It is precisely at ultra-short distances – ranging from a few centimeters to a few meters – that classic optical and microwave sensors lose their effectiveness, leaving the vehicle unprotected before low or barely noticeable obstacles. Under these specific conditions, ultrasonic sensors remain the only alternative tool that provides millimeter measurement accuracy using acoustic echolocation. The operation of these sensors ensures the safe execution of complex maneuvers in critically restricted spaces where cameras are unable to correctly evaluate the geometry of objects.

Quick Take

Ultrasonic sensors provide millimeter accuracy at ultra-short distances.
Sensors calculate distance based on the time-of-flight of a sound pulse.
Creating a high-quality parking sensor dataset includes capturing the exact distance, filtering background noise, detecting vacant spaces, and creating ground truths for low objects.
At the data preparation stage, ultrasonic signals are time-aligned with the video stream of cameras and LiDAR point clouds down to milliseconds to build a unified map of the environment.
The low-speed ADAS annotation process requires maximum detailization of chaotic scenarios, as a calculation error of just a few centimeters leads to an accident.

How Ultrasonic Sensors Support Parking AI

Principle of Sensor Operation

An ultrasonic sensor operates on the principle of echolocation, very similar to how bats or dolphins orient themselves in space. The sensor, mounted on the car's bumper, emits a short sound pulse at a high frequency that is inaudible to the human ear. This acoustic wave travels through the air, reflects off the nearest object, and returns in the form of an echo. The sensor captures the exact time it took for the sound to travel there and back, and based on this, the onboard computer instantly calculates the physical space around the car.

In order to teach artificial intelligence to correctly interpret these sound echoes, developers need a high-quality parking sensor dataset. During the markup of such a dataset, specialists help the AI clearly determine several important parameters:

Exact distance. Detailed obstacle distance labeling is conducted, thanks to which the system understands the distance to an adjacent car or wall with millimeter precision.
Presence and type of obstacles. The AI learns to distinguish real threats (pillars, fences, trees) from random noise, such as heavy rain, exhaust gases, or leaves on the asphalt.
Proximity to curbs. Special curb detection ground truths are created, which help the car see low obstacles under the wheels and protect rims and sills from damage.
Vacant space. The robot calculates the geometry of an empty space in a parking lot, checking whether the car fits there by its dimensions.

The Role of Ultrasound in Driver Assistance Systems

The collected and correctly marked data becomes the foundation for training driver assistance systems at low speeds. Because the car moves slowly during parking or maneuvering in yards, cameras often cannot adequately evaluate the volume and distance to objects in the immediate proximity zone. Ultrasound closes this technical gap, ensuring the stable operation of three key safety systems:

System Name	Role of Ultrasonic Sensors	Main Benefit for Safety
Parking Assist	Continuously scans the space during the driver's backward or forward movement, warning with acoustic signals about approaching a wall or another car.	Prevents minor household collisions and "scrapes" in tight yards or underground parking lots.
Automated Parking	Acts as the main sensory organ for the autopilot: it searches for a vacant pocket among cars and guides the vehicle along an ideal trajectory.	The car can park in a narrow space completely autonomously without the risk of clipping adjacent vehicles.
Low-Speed ADAS	Monitors blind spots around bumpers during startup from a standstill or crawling in traffic jams, initiating emergency braking if necessary.	Protects against colliding with sudden obstacles, low pillars, or pets that the driver cannot see from behind the hood.

Thanks to engineers preparing high-quality automated parking training data, modern AI can combine information from all ultrasonic sensors simultaneously. This allows the car to "see" the environment invisible to the driver's eyes and guarantees that automatic systems will trigger in time and flawlessly.

Building the Database

The creation of reliable artificial intelligence for auto-parking is impossible without long-term and systematic work on data preparation. Central to this process is the formation of a parking sensor dataset – a specialized dataset that collects thousands of hours of acoustic signal recordings in real city conditions. Only after passing a full cycle from technical collection to multi-level verification do these raw files turn into fuel for training parking neural networks.

Data Preparation Life Cycle

The process of transforming ultrasonic echo signals into a finished commercial product is strictly sequential and consists of five key stages:

Data Collection. Specially equipped laboratory vehicles drive hundreds of kilometers, capturing millions of ultrasonic pulses from various objects on roads and parking lots.
Sensor Synchronization. Since ultrasound works in tandem with surround-view cameras and LiDARs, engineers align all data streams in time with millisecond precision so that each sound echo exactly matches the visual image of the obstacle.
Annotation. Annotation specialists conduct a complex ultrasonic annotation process, cleaning acoustic waves of background noise and marking precise physical boundaries around found objects.
Quality Validation. Finished information packages undergo strict automatic and manual control, where the accuracy of labels, the absence of missed obstacles, and the correctness of distance measurements are verified.
Model Training. Validated and cleaned data are handed over to machine learning engineers, who use this array as automated parking training data to train the final autopilot algorithms.

Necessary Diversity of Scenarios

To ensure the vehicle's onboard computer does not get lost in real life, a high-quality dataset must contain a huge number of diverse and complex spatial situations. AI models must clearly process the following scenarios:

Parallel Parking. Teaches the system to precisely determine the boundaries between the bumpers of cars already parked along the curb and to react to the corners of other people's cars in time.
Perpendicular Parking. Helps the robot see the sides of adjacent vehicles and maintain an even distance on both sides when entering a pocket.
Narrow Spaces. Trains the AI for millimeter precision in conditions where sensors from all sides simultaneously capture critical proximity to walls, pillars, or pedestrian zones.
Garages. A scenario where ultrasound encounters a strong reverberation effect from close closed surfaces, and the model learns to distinguish real walls from false signals.
Underground Parking Lots. Complex zones with a large number of concrete columns, ramps, boundary chains, and specific floor coatings that require maximum concentration from the echolocation system.

Specifics of Low-Speed ADAS Annotation

Driving automation at low speeds places unprecedented demands on the accuracy of spatial perception by artificial intelligence. While on a highway, it is sufficient for active safety systems to capture the general dimensions of a traffic lane and large objects at a distance of tens of meters; during parking or maneuvering, the count goes to millimeters. Any error in calculations or a signal delay of just a few centimeters instantly leads to financial losses due to damage to the body, rims, or surrounding infrastructure. The very essence of the low-speed ADAS annotation process is to teach the AI to see micro-details and instantly react to hidden threats.

Critical Scenarios for Low-Speed Markup

To create reliable driver assistance algorithms, annotator-engineers carefully mark specific road situations where ultrasonic sensors operate at the limit of their capabilities:

Traffic Jams. A scenario in which the car is constantly surrounded by other vehicles, motorcyclists, and pedestrians. The AI must clearly distinguish the distance to the bumper of the car ahead, while ignoring heat flows from exhaust pipes and vibrations from heavy machinery.
City Yards. A chaotic environment with a high level of unpredictability. Markup helps the model detect low trash cans, open manholes, chaotically parked cars, as well as capture the sudden appearance of children or pets in the bumper's blind spot zones in time.
Parking Lots. Zones with a high concentration of static obstacles of various shapes. Here, complex reflections of sound from concrete columns, ramps, barriers, and metal fences are annotated so that the car can flawlessly turn around in restricted pockets.
Narrow Passages. Specific locations such as historic streets, building arches, or road construction zones. In such conditions, sensors on both sides of the bumper simultaneously capture critical proximity to walls, and the AI learns to guide the car exactly down the center of the available corridor.
Maneuvering Between Cars. The process of passing oncoming traffic in tight alleys or exiting a dense parking row. Markup trains the neural network to precisely evaluate the dynamic change in distance to other people's mirrors and fenders directly while turning the steering wheel.

The Cost of an Error in Ultrasonic Markup

Since an ultrasonic sensor emits a wave in the shape of a cone, the raw signal reflects from objects at various angles. The annotator's task is to capture the presence of an echo and clearly separate the useful signal from interference. If a marker mistakenly labels the echo from tall grass or uneven asphalt as a threat, the car will constantly perform dangerous emergency braking in the middle of an empty road.

Conversely, if a barely noticeable echo from a low reinforced concrete pillar or rebar is missed, the parking autopilot will simply run over the obstacle. That is why high-quality micro-markup and filtering of echolocation data are the main safeguard that allows modern cars to safely take full control during complex maneuvers in the urban jungle.

FAQ

How do annotators visualize raw ultrasonic signals to conduct markup on a computer screen?

Since sound waves cannot be seen directly, specialized software transforms acoustic signals into graphical amplitude-time diagrams or spectrograms. Markup engineers see graphs of energy spikes from the reflected echo, where each axis displays signal delay time and signal strength. Most often, this graphical information is overlaid in the form of colored cones or point clouds directly onto the 3D scene obtained from the car's synchronized cameras.

How do weather conditions, such as severe frost or heat, affect ultrasonic data and the process of their annotation?

The speed of sound propagation in air is not constant and directly depends on the temperature and humidity of the surrounding environment. During severe temperature fluctuations, raw sensors without calibration begin to output false distances to objects. During annotation, engineers implement mathematical compensation coefficients, adding precise air temperature indicators to the dataset metadata so that the AI model learns to correct distance calculations in any weather.

How do AI algorithms trained on ultrasonic datasets interact with pedestrians compared to classic cameras?

Cameras recognize a pedestrian by their visual silhouette and can determine the direction of their gaze or movement, but they work worse in complete darkness or fog. Ultrasonic sensors do not know that it is a person in front of them, but they instantly capture the appearance of any physical body near the bumper, regardless of lighting. Thanks to the markup of dynamic obstacles, the parking AI is capable of capturing a sharp reduction in distance to a pedestrian's legs in milliseconds and activating the brakes even before the driver orients themselves via mirrors.

Are there specific object materials that reflect ultrasound poorly, and how are they marked in a "parking sensor dataset"?

Materials with a soft or porous texture, such as thick clothing of pedestrians, animal fur, foam rubber, or bushes, intensively absorb sound waves instead of reflecting them. Such objects return a very weak, decaying echo that the AI can easily confuse with ordinary noise. In the annotation process, such specific decaying signals are marked with a special class "absorbing obstacle" to teach the model to be maximally sensitive to weak amplitude spikes.

The blind zone of an ultrasonic sensor occurs directly near its surface because the sensor cannot simultaneously emit a wave and receive its echo. During the development of automated parking, this shortcoming is compensated for by maintaining the movement history of the vehicle. In annotated datasets, the trajectory is sequentially marked: if an object was captured at a distance of 30 cm and then disappeared during approach, the AI model remembers its virtual presence in the blind zone and keeps the brakes activated.

How are noises from different types of asphalt surfaces (gravel, cobblestones, smooth concrete) filtered during the markup process?

Rough gravel or historic cobblestones create a huge number of chaotic, small reflections of sound, which look like solid information noise called ground clutter on a spectrogram. Experienced annotators use frequency filtering algorithms to separate this low-amplitude background hum from the clear, strong signals of real obstacles. The AI model is trained on many types of surfaces so that it can automatically cut off "noise under the wheels" without reducing the overall sensitivity of the system.

In what way are moving obstacles (for example, another car backing up) marked in a dataset for low-speed ADAS?

To mark dynamic objects in low-speed ADAS annotation, frame-by-frame markup of sequences is used. Each acoustic pulse is marked as part of a temporal track where the Doppler effect – the change in frequency of the reflected wave depending on the closing speed of objects – is captured. This allows the parking AI to instantly calculate whether an object is stationary or moving toward the car, and to forecast the time until a potential collision.

What automation tools are used to accelerate the markup process for large volumes of ultrasonic data?

To speed up the work, engineers use automatic pre-markup algorithms built on the basis of geometric triangulation and data from LiDAR. The computer independently calculates the approximate distance to objects based on the lidar point cloud and places primary markers on ultrasonic graphs. The human annotator plays the role of a validator in this process: they only correct automation errors, clean complex echo anomalies, and confirm the accuracy of the final data package.