Parquery AI - What makes our Smart Parking solution outstanding?

 

Why does Parquery’s solution work for any vehicle, parking situation, and weather condition?

Exploring the strengths of Parquery’s powerful Artificial Intelligence algorithms

 

Parquery’s smart parking solution uses cameras and Artificial Intelligence (AI) to monitor parking lots. The intelligent algorithms run either in the cloud or locally on the parking premises and detect vehicles in the camera images and determine whether individual parking spots are occupied.

Parquery’s AI solution is robust even in adverse conditions

Footage from one of Parquery’s customers demonstrates how robust Parquery’s AI solution is even in adverse conditions with several factors present simultaneously: at night time with little light, heavy rain, intense glare and reflections, and in the presence of occlusions.

Learn more about Parquery's project with SBB

 

Learn and adapt

 

The power of Parquery’s solution lies in its advanced Deep Neural Network algorithms - machine learning algorithms that allow the system to adapt to:

Any vehicle or any object

 

e.g., cars, trucks, buses, trains, motorbikes, bicycles, boats, tractors, construction or mining machinery, planes

Parquery detects boats in marinas

Boats

Parquery detects planes in airports

Planes

Parquery detects trucks on highways

Trucks

Parquery detects people and vehicles on construction sites

Construction machinery

Parquery detects pedestrians in public and private areas

Persons or silhouettes

Any parking situation

 

e.g., indoors or outdoors, roadside parking, multi-storey car parks, highway truck parking, traffic counting, fuel stations, electric car charging

All Parquery needs is images from cameras, even fisheye cameras

Indoors

Parquery detects vehicles outdoors with any camera

Outdoors

Parquery detects vehicles on the roadside

Roadside

Parquery detects which fuel pump is available

Fuel stations

Parquery displays highway parking availability on rest areas

Highway truck parking

Any environmental condition

Any weather

e.g., glaring sunshine, glistening rain, covering snow, fog

Parquery detects vehicles even in glaring sun

Glaring sun

Parquery detects vehicles even in rain

Reflecting tarmac in the rain

Parquery detects vehicles even in fog

Fog at night

Parquery detects vehicles even in snow

Snow

 

Any lighting

At day and night, with visible light or infrared, natural or artificial light from street lamps or spotlights, uniform illumination and extreme contrasts (e.g. strong sunlight), moving shadows

Parquery detects vehicles even at night

Night

Parquery detects vehicles even at twilight

Twilight: natural and artificial light

Parquery detects vehicles even with strong contrasts

Strong contrasts

Parquery detects vehicles even with moving shadows

Moving shadows

Any camera and any lens

e.g., CCTVs, webcams, steady and rotating, high-resolution and low, fisheye to narrow-angle lenses

All Parquery needs is images from cameras, even with wide-angles

Wide-angle lens

All Parquery needs is images from cameras, even fisheye cameras

Fisheye lens

All Parquery needs is images from cameras

Regular lens

Training: Learning by doing

 

Parquery’s system adjusts to a new environment by training or experience - that means it learns from its mistakes - just like us. Starting with a generic model generated from various parking situations, this model is refined as time passes and optimized for the specific characteristics through training itself on the new data it receives. Thus, it ​​gains experience and increases the probability of a correct outcome next time.

For example, the AI might incorrectly decide that a parked flatbed truck is a vacant parking spot due to the uniform appearance of its cargo deck, overlooking another differentiator that did not seem relevant for the decision. Through feedback about this mistake and training, the AI algorithm identifies and learns a vital differentiator to distinguish between an empty parking bay and an empty cargo deck. Thus it improves the probability of a correct differentiation subsequently.

Beyond the evident differences between parking lots - such as layout, arrangement of the spots, obstructions, or whether indoors or outdoors - other characteristics include lighting and weather conditions, the types of vehicles encountered, and visual distortions due to camera optics.

 

Empty-trailer
Empty_trailer
Empty-trailer-flatbed

Before: trailer and flatbed truck incorrectly classified as vacant spaces.

Note that the appearance of the loading decks is similar to the road surface: the patterns on the loading deck have the mottled appearance of patched asphalt, tree shadows, oil stains, or puddles after rain.

 

Parquery trains its software to adapt to any vehicle, such as flatbeds
Parquery trains its software to adapt to any vehicle, such as empty trailers
Parquery trains its software to adapt to any vehicle, such as trailers
Parquery trains its software to adapt to any vehicle, even flatbeds

After retraining: the system correctly detects flatbed trucks without features, irrespective of where they are parked and differences in the camera perspective.

From good to great

 

Typically, the out-of-the-box accuracy of Parquery’s learning solution is about 95% before adapting to the specific parking lot. The 5% error concerns undetected vehicles (so-called false negatives) and empty parking bays erroneously interpreted as occupied (false positives). Throughout the first month, Parquery collects and annotates image data of the parking lot to create the learning material for the AI algorithms.

By the end of the first month, Parquery delivers 99% accuracy thanks to training and adapting the algorithms to the requirements of the parking situation at hand.

The real world is no mathematical model

 

The physical world is varied and does not conform to rigid norms, abstract rules, and theoretical models. Parquery’s system is built for this world as it is: This means it is robust to and accounts for imperfections, deviations, oversights, or changes.

Drivers might not conform to the boundary markings, park across or diagonally over several spots. Vehicles might be too large to fit into a single space but stick out or use several parking bays. Tree crowns, overhanging roofs, or large trucks might partially hide vehicles from the camera view. Cameras might only partially capture parking spaces at the edge of their fields of view.

Parquery detects vehicles even if they are parked across
Parquery detects vehicles even if they are not inside marked areas
Parquery detects vehicles even if they are parked diagonally
Parquery detects vehicles even if they are wrongly parked
Parquery detects vehicles even if they are parked slightly askew
Parquery detects vehicles even if they are diagonal to the spot
Parquery detects vehicles even if they are parked over two spots
Parquery detects vehicles even if they are not parked fully in the spot

Parquery's algorithms detect vehicles regardless of whether they are parked improperly, at an angle, across, or protrude beyond the boundaries.

Parquery detects vehicles even if they are parked across two spots
Parquery detects vehicles even if they are parked over three spots
Parquery detects vehicles even if the flatbed is across two spots
Parquery detects vehicles even if trucks are across multiple spots
Parquery detects vehicles even if they are parked over four spaces

Even if the vehicles are parked across several parking spaces and do not adhere to the specified marking lines at all, Parquery's algorithms give accurate results.

Never mind occlusions and imperfections

 

For Parquery’s system to work, neither camera coverage nor parking habits have to be perfect. The algorithms function with incomplete data and detect partial vehicles - whether somewhat hidden by others, by pedestrians or bikes passing through, parked over several spots, being loaded or unloaded, or at the edge of the camera image where only a portion is visible.

Parquery detects vehicles even with tree occlusions
Parquery detects vehicles even with slight tree occlusions
Parquery detects vehicles even with lamp occlusions
Parquery detects vehicles even with occlusions from billboards
Parquery detects vehicles even behind construction fences

Permanent (trees, lamp posts, billboards, street signs, roofs) or temporary occlusions (no stop signs, construction site fencing)

Parquery detects vehicles even behind passing buses
Parquery detects vehicles even behind large passing buses
Parquery's software can differentiate between bikes, people and vehicles
Parquery detects vehicles even behind passing cyclists
Parquery detects vehicles even with doors opened
Parquery detects vehicles even with people around
Parquery detects vehicles even when loading in the box

Partial occlusions such as passing busses, coaches, trucks, pedestrians, or bikes, and changes to appearance caused by e.g., open doors when loading and unloading or shopping trolleys do not impair detection.

Parquery does not buckle even under added strain

 

At the same time, Parquery’s algorithms do the tightrope walk to remain robust to clutter, noise and background patterns that can resemble vehicle parts taken out of context. Oil stains, shadows, reflections, broken road surfaces, patches of repaired tarmac, tire marks, shopping trolleys, trash, or litter left behind are typical such issues.

 

Parquery differentiates people from vehicles
Parquery differentiates random objects on the ground from vehicles
Parquery differentiates trolleys from vehicles
Parquery differentiates shopping carts from vehicles
Parquery differentiates cyclists from vehicles
Parquery differentiates passing cyclists from vehicles

Parquery’s algorithms differentiate pedestrians, bikes, suitcases, trolleys, cyclists, etc from vehicles.

 

Parquery’s algorithms are robust to rubbish on the ground
Parquery differentiates oil stains from vehicles
Parquery differentiates stains and tyre marks from vehicles
Parquery differentiates autumn leaves from vehicles
Parquery detects vehicles even under heavy snow

Litter, oil and water stains, drying rain, tyre marks, autumn leaves, or snow do not mislead Parquery’s AI.

 

Parquery’s algorithms are robust from rain drops
Parquery’s algorithms are robust to condensation
Parquery’s algorithms are robust to image noise, such as spider webs
Parquery’s algorithms are robust to spider webs in glaring sun
Parquery’s algorithms are robust to noise from lens flare

Parquery is also robust to other typical issues with camera-based systems, e.g., lens flare in backlight, scratches and pollutants on the camera lens, such as dust, rain drops, condensation, or even cobwebs that catch and reflect the light.

 

What are the fundamental components for Parquery's solution?

Even the best AI capitulates at times

 

 

No system is 100% fail-safe - if the camera view is completely blocked, even the cleverest AI cannot make up for it.

Note, however, that in the original footage shown, one car (yellow dot) is still detected even though it is barely visible.

 

If cameras have a blocked view, no detection is possible

More than 80 projects in over 30 countries worldwide

 

Explore our references