The first necessary condition to support autonomous decision making is to be able to distinguish drivable from non-drivable portions of the immediate driving environment. Free space mapping is the basis for navigation, path planning, and obstacle avoidance, which require a reliable estimate of the empty vs. occupied space in the vehicle’s environment.
Free space mapping used for vehicle autonomy cannot be performed exclusively with one sensor. The purpose of sensor redundancy is not only to provide a fail-safe mechanism, but also to exploit the individual sensors’ advantages to achieve better coverage. An imaging radar, with its high sensitivity, high-resolution, full spatial sensing (including elevation), and all-weather all-visibility performance, is the perfect candidate to counter the contrast-based free space mapping performed by cameras.
Even though cameras provide adequate azimuth and elevation mapping and are a commonly used sensor for free space mapping, a lack of reliable range resolution or sufficient depth perception results in their inability to accurately measure frontal target distance, a crucial requirement for path planning. Radar, on the other hand, provides highly accurate target distance mapping and is operational in all weather and lighting conditions making it the ideal (and only necessary) complementary sensor to a camera.
LiDAR provides a satisfactory free space solution for mid ranges. Lidar does not measure Doppler directly and must rely on tracking to determine which objects are stationary. Some LiDAR offers ultra-high resolution of 0.1 degree, higher than radar’s 1 degree; however, this does not provide a meaningful impact for free space mapping since driving decisions require wide safety margins and cannot be executed on 0.1 degree separations. Taking into account LiDAR costs and limited functionality in challenging weather conditions, it is better suited as a backup sensor for an imaging radar and camera suite.
In contrast to the aforementioned 0.1 degree resolution of LiDAR, traditional radar with 5 degree (or worse) resolution are also inadequate for free-space mapping as they cannot provide a detailed enough occupancy grid of their driving scenario. In low resolution radars, objects appear to be much bigger than they actually are, making it impossible to understand their boundaries and exact locations.
Due to their low azimuth resolution and lack of elevation resolution, traditional radars that were originally designed for controlling emergency braking and adaptive cruise control are limited to detecting moving vehicles and typically discard all detections of the stationary environment as clutter. While moving objects can be separated by their Doppler measurements, stationary objects may be unified by mistake into one oversized block, or be discarded.
Arbe’s Imaging Radar, on the other hand, is very effective in detecting stationary objects, determining road geometry and curvature while mapping objects that need to be avoided such as guardrails, concrete walls, vehicles parked in-lane, etc.
Traditional radars have a high level of false alarms. False alarms trigger radar to report phantom objects and false positives. For this reason, traditional radar-based post-processing algorithms raise the detection threshold, which results in decreased radar sensitivity and may lead to hazardous situations or even accidents.
Superior channel counts that yield improved spatial separation in imaging, such as Arbe’s radar, together with advanced post-processing reduce false alarms with close to zero instances of phantom objects, eliminating both false-positive and false-negative scenarios. This enables the generation of a free-space map the vehicle can rely on.
A free space grid map is a 2D occupancy map whose interpretation is a “pixel”-wise probability of being free. This map takes into account current frame information as well as a “memory-component” from previous frames to increase the confidence level of free-space inference, ensuring spurious, false positive detections that are detected despite other noise cleaning algorithms are washed out, and correcting for transient occlusions and multipath reflections.
As mentioned, the free space map is a 2D map, offering a bird’s eye view of the driving environment. In radar terms, this is a range-azimuth map. Each pixel on the map tells us if the corresponding range and azimuth are drivable or not. An important subtlety, however, is that being a 2D output map does not mean that the third spatial dimension, elevation, bears no significance for the ability to reliably map the free space that surrounds the car. In fact, the opposite is true; being able to tell the difference between an obstacle that is irrelevant for driving, such as a road sign, and an obstacle that should be avoided, such as construction-barriers, lies in elevation resolution, which makes imaging radar crucial for free space mapping.
Imaging radar satisfies key requirements for free space mapping, namely high spatial resolution in all dimensions regardless of lighting or weather conditions, at long range with high reliability. Additionally, its complementary high Doppler resolution allows it to distinguish efficiently between stationary and dynamic detections. Together these capabilities make imaging radar a mandatory sensor suite component to map free space in real time.
Connect to learn more
