This sensor registers an event each time its parent actor collisions against something in the world. Several collisions may be detected during a single simulation step.
To ensure that collisions with any kind of object are detected, the server creates "fake" actors for elements such as buildings or bushes so the semantic tag can be retrieved to identify it.
The camera provides a raw data of the scene codifying the distance of each pixel to the camera (also known as **depth buffer** or **z-buffer**) to create a depth map of the elements.
The image codifies depth value per pixel using 3 channels of the RGB color space, from less to more significant bytes: _R -> G -> B_. The actual distance in meters can be
The output [carla.Image](python_api.md#carla.Image) should then be saved to disk using a [carla.colorConverter](python_api.md#carla.ColorConverter) that will turn the distance stored in RGB channels into a __[0,1]__ float containing the distance and then translate this to grayscale.
There are two options in [carla.colorConverter](python_api.md#carla.ColorConverter) to get a depth view: __Depth__ and __Logaritmic depth__. The precision is milimetric in both, but the logarithmic approach provides better results for closer objects.
Reports current [gnss position](https://www.gsa.europa.eu/european-gnss/what-gnss) of its parent object. This is calculated by adding the metric position to an initial geo reference location defined within the OpenDRIVE map definition.
Provides measures that accelerometer, gyroscope and compass would retrieve for the parent object. The data is collected from the object's current state.
Registers an event each time its parent crosses a lane marking.
The sensor uses road data provided by the OpenDRIVE description of the map to determine whether the parent vehicle is invading another lane by considering the space between wheels.
However there are some things to be taken into consideration:
* Discrepancies between the OpenDRIVE file and the map will create irregularities such as crossing lanes that are not visible in the map.
* The output retrieves a list of crossed lane markings: the computation is done in OpenDRIVE and considering the whole space between the four wheels as a whole. Thus, there may be more than one lane being crossed at the same time.
This sensor simulates a rotating LIDAR implemented using ray-casting.
The points are computed by adding a laser for each channel distributed in the vertical FOV. The rotation is simulated computing the horizontal angle that the Lidar rotated in a frame. The point cloud is calculated by doing a ray-cast for each laser in every step.
A LIDAR measurement contains a package with all the points generated during a `1/FPS` interval. During this interval the physics are not updated so all the points in a measurement reflect the same "static picture" of the scene.
This output contains a cloud of simulation points and thus, it can be iterated to retrieve a list of their [`carla.Location`](python_api.md#carla.Location):
The information of the LIDAR measurement is enconded 4D points. Being the first three, the space points in xyz coordinates and the last one intensity loss during the travel. This intensity is computed by the following formula.
`a` — Attenuation coefficient. This may depend on the sensor's wavelenght, and the conditions of the atmosphere. It can be modified with the LIDAR attribute `atmosphere_attenuation_rate`.
`d` — Distance from the hit point to the sensor.
For a better realism, points in the cloud can be dropped off. This is an easy way to simulate loss due to external perturbations. This can done combining two different.
*__General drop-off__ — Proportion of points that are dropped off randomly. This is done before the tracing, meaning the points being dropped are not calculated, and therefore improves the performance. If `dropoff_general_rate = 0.5`, half of the points will be dropped.
*__Instensity-based drop-off__ — For each point detected, and extra drop-off is performed with a probability based in the computed intensity. This probability is determined by two parameters. `dropoff_zero_intensity` is the probability of points with zero intensity to be dropped. `dropoff_intensity_limit` is a threshold intensity above which no points will be dropped. The probability of a point within the range to be dropped is a linear proportion based on these two parameters.
Additionally, the `noise_stddev` attribute makes for a noise model to simulate unexpected deviations that appear in real-life sensors. For positive values, each point is randomly perturbed along the vector of the laser ray. The result is a LIDAR sensor with perfect angular positioning, but noisy distance measurement.
The rotation of the LIDAR can be tuned to cover a specific angle on every simulation step (using a [fixed time-step](adv_synchrony_timestep.md)). For example, to rotate once per step (full circle output, as in the picture below), the rotation frequency and the simulated FPS should be equal. <br>__1.__ Set the sensor's frequency `sensors_bp['lidar'][0].set_attribute('rotation_frequency','10')`. <br>__2.__ Run the simulation using `python config.py --fps=10`.
To ensure that collisions with any kind of object are detected, the server creates "fake" actors for elements such as buildings or bushes so the semantic tag can be retrieved to identify it.
The sensor creates a conic view that is translated to a 2D point map of the elements in sight and their speed regarding the sensor. This can be used to shape elements and evaluate their movement and direction. Due to the use of polar coordinates, the points will concentrate around the center of the view.
Points measured are contained in [carla.RadarMeasurement](python_api.md#carla.RadarMeasurement) as an array of [carla.RadarDetection](python_api.md#carla.RadarDetection), which specifies their polar coordinates, distance and velocity.
The provided script `manual_control.py` uses this sensor to show the points being detected and paint them white when static, red when moving towards the object and blue when moving away:
<td>Opening of the camera lens. Aperture is <code>1/fstop</code> with typical lens going down to f/1.2 (larger opening). Larger numbers will reduce the Depth of Field effect.</td>
<td>Can be <code>manual</code> or <code>histogram</code>. More in <ahref="https://docs.unrealengine.com/en-US/Engine/Rendering/PostProcessEffects/AutomaticExposure/index.html">UE4 docs</a>.</td>
<td>Logarithmic adjustment for the exposure. 0: no adjustment, -1:2x darker, -2:4 darker, 1:2x brighter, 2:4x brighter.</td>
<tr>
<td><code>exposure_min_bright</code></td>
<td>float</td>
<td>0.1</td>
<td>In <code>exposure_mode: "histogram"</code>. Minimum brightness for auto exposure. The lowest the eye can adapt within. Must be greater than 0 and less than or equal to <code>exposure_max_bright</code>.</td>
<tr>
<td><code>exposure_max_bright</code></td>
<td>float</td>
<td>2.0</td>
<td>In `exposure_mode: "histogram"`. Maximum brightness for auto exposure. The highestthe eye can adapt within. Must be greater than 0 and greater than or equal to `exposure_min_bright`.</td>
<tr>
<td><code>exposure_speed_up</code></td>
<td>float</td>
<td>3.0</td>
<td>In <code>exposure_mode: "histogram"</code>. Speed at which the adaptation occurs from dark to bright environment.</td>
<tr>
<td><code>exposure_speed_down</code></td>
<td>float</td>
<td>1.0</td>
<td>In <code>exposure_mode: "histogram"</code>. Speed at which the adaptation occurs from bright to dark environment.</td>
<tr>
<td><code>calibration_constant</code></td>
<td>float</td>
<td>16.0</td>
<td>Calibration constant for 18% albedo.</td>
<tr>
<td><code>focal_distance</code></td>
<td>float</td>
<td>1000.0</td>
<td>Distance at which the depth of field effect should be sharp. Measured in cm (UE units).</td>
<tr>
<td><code>blur_amount</code></td>
<td>float</td>
<td>1.0</td>
<td>Strength/intensity of motion blur.</td>
<tr>
<td><code>blur_radius</code></td>
<td>float</td>
<td>0.0</td>
<td>Radius in pixels at 1080p resolution to emulate atmospheric scattering according to distance from camera.</td>
<tr>
<td><code>motion_blur_intensity</code></td>
<td>float</td>
<td>0.45</td>
<td>Strength of motion blur [0,1].</td>
<tr>
<td><code>motion_blur_max_distortion</code></td>
<td>float</td>
<td>0.35</td>
<td>Max distortion caused by motion blur. Percentage of screen width.</td>
<td>Percentage of screen width objects must have for motion blur, lower value means less draw calls.</td>
<tr>
<td><code>slope</code></td>
<td>float</td>
<td>0.88</td>
<td>Steepness of the S-curve for the tonemapper. Larger values make the slope steeper (darker) [0.0, 1.0].</td>
<tr>
<td><code>toe</code></td>
<td>float</td>
<td>0.55</td>
<td>Adjusts dark color in the tonemapper [0.0, 1.0].</td>
<tr>
<td><code>shoulder</code></td>
<td>float</td>
<td>0.26</td>
<td>Adjusts bright color in the tonemapper [0.0, 1.0].</td>
<tr>
<td><code>black_clip</code></td>
<td>float</td>
<td>0.0</td>
<td>This should NOT be adjusted. Sets where the crossover happens and black tones start to cut off their value [0.0, 1.0].</td>
<tr>
<td><code>white_clip</code></td>
<td>float</td>
<td>0.04</td>
<td>Set where the crossover happens and white tones start to cut off their value. Subtle change in most cases [0.0, 1.0].</td>
<tr>
<td><code>temp</code></td>
<td>float</td>
<td>6500.0</td>
<td>White balance in relation to the temperature of the light in the scene. <b>White light:</b> when this matches light temperature. <b>Warm light:</b> When higher than the light in the scene, it is a yellowish color. <b>Cool light:</b> When lower than the light. Blueish color.</td>
<tr>
<td><code>tint</code></td>
<td>float</td>
<td>0.0</td>
<td>White balance temperature tint. Adjusts cyan and magenta color ranges. This should be used along with the white balance Temp property to get accurate colors. Under some light temperatures, the colors may appear to be more yellow or blue. This can be used to balance the resulting color to look more natural.</td>
This sensor integrates the [C++ Library for Responsibility Sensitive Safety](https://github.com/intel/ad-rss-lib) in CARLA. It is disabled by default in CARLA, and it has to be explicitly built in order to be used.
The RSS sensor calculates the RSS state of a vehicle and retrieves the current RSS Response as sensor data. The [carla.RssRestrictor](python_api.md#carla.RssRestrictor) will use this data to adapt a [carla.VehicleControl](python_api.md#carla.VehicleControl) before applying it to a vehicle.
These controllers can be generated by an *Automated Driving* stack or user input. For instance, hereunder there is a fragment of code from `PythonAPI/examples/manual_control_rss.py`, where the user input is modified using RSS when necessary.
The blueprint for this sensor has no modifiable attributes. However, the [carla.RssSensor](python_api.md#carla.RssSensor) object that it instantiates has attributes and methods that are detailed in the Python API reference. Here is a summary of them.
<td>Enables/Disables the <ahref="https://intel.github.io/ad-rss-lib/ad_rss_map_integration/HandleRoadBoundaries">stay on road</a> feature. Default is <b>Off</b>.</td>
This sensor works fully on the client side. There is no blueprint in the server. Changes on the attributes will have effect __after__ the *listen()* has been called.
The methods available in this class are related to the routing of the vehicle. RSS calculations are always based on a route of the ego vehicle through the road network.
The sensor allows to control the considered route by providing some key points, which could be the [carla.Transform](python_api.md#carla.Transform) in a [carla.Waypoint](python_api.md#carla.Waypoint). These points are best selected after the intersections to force the route to take the desired turn.
* __Output:__ [carla.SemanticLidarMeasurement](python_api.md#carla.SemanticLidarMeasurement) per step (unless `sensor_tick` says otherwise).
This sensor simulates a rotating LIDAR implemented using ray-casting that exposes all the information about the raycast hit. Its behaviour is quite similar to the [LIDAR sensor](#lidar-sensor), but there are two main differences between them.
* The raw data retrieved by the semantic LIDAR includes more data per point.
* Coordinates of the point (as the normal LIDAR does).
* The cosine between the angle of incidence and the normal of the surface hit.
* Instance and semantic ground-truth. Basically the index of the CARLA object hit, and its semantic tag.
* The semantic LIDAR does not include neither intensity, drop-off nor noise model attributes.
The points are computed by adding a laser for each channel distributed in the vertical FOV. The rotation is simulated computing the horizontal angle that the LIDAR rotated in a frame. The point cloud is calculated by doing a ray-cast for each laser in every step.
A LIDAR measurement contains a package with all the points generated during a `1/FPS` interval. During this interval the physics are not updated so all the points in a measurement reflect the same "static picture" of the scene.
This output contains a cloud of lidar semantic detections and therefore, it can be iterated to retrieve a list of their [`carla.SemanticLidarDetection`](python_api.md#carla.SemanticLidarDetection):
```py
for detection in semantic_lidar_measurement:
print(detection)
```
The rotation of the LIDAR can be tuned to cover a specific angle on every simulation step (using a [fixed time-step](adv_synchrony_timestep.md)). For example, to rotate once per step (full circle output, as in the picture below), the rotation frequency and the simulated FPS should be equal. <br>
__1.__ Set the sensor's frequency `sensors_bp['lidar'][0].set_attribute('rotation_frequency','10')`. <br>
__2.__ Run the simulation using `python config.py --fps=10`.
This camera classifies every object in sight by displaying it in a different color according to its tags (e.g., pedestrians in a different color than vehicles).
When the simulation starts, every element in scene is created with a tag. So it happens when an actor is spawned. The objects are classified by their relative file path in the project. For example, meshes stored in `Unreal/CarlaUE4/Content/Static/Pedestrians` are tagged as `Pedestrian`.
The server provides an image with the tag information __encoded in the red channel__: A pixel with a red value of `x` belongs to an object with tag `x`.
This raw [carla.Image](python_api.md#carla.Image) can be stored and converted it with the help of __CityScapesPalette__ in [carla.ColorConverter](python_api.md#carla.ColorConverter) to apply the tags information and show picture with the semantic segmentation.
It requires some C++ coding. Add a new label to the `ECityObjectLabel` enum in "Tagger.h", and its corresponding filepath check inside `GetLabelByFolderName()` function in "Tagger.cpp".
A Dynamic Vision Sensor (DVS) or Event camera is a sensor that works radically differently from a conventional camera. Instead of capturing
intensity images at a fixed rate, event cameras measure changes of intensity asynchronously, in the form of a stream of events, which encode per-pixel
brightness changes. Event cameras possess outstanding properties when compared to standard cameras. They have a very high dynamic range (140 dB
versus 60 dB), no motion blur, and high temporal resolution (in the order of microseconds). Event cameras are thus sensors that can provide high-quality
visual information even in challenging high-speed scenarios and high dynamic range environments, enabling new application domains for vision-based
The provided script `manual_control.py` uses the DVS camera in order to show how to configure the sensor, how to get the stream of events and how to depict such events in an image format, usually called event frame.
DVS is a camera and therefore has all the attributes available in the RGB camera. Nevertheless, there are few attributes exclusive to the working principle of an Event camera.
<td> Refractory period (time during which a pixel cannot fire events just after it fired one), in nanoseconds. It limits the highest frequency of triggering events.</td>
<tr>
<td><code>use_log</code></td>
<td>bool</td>
<td>true</td>
<td>Whether to work in the logarithmic intensity scale.</td>
<td>Epsilon value used to convert images to log: <code>L = log(eps + I / 255.0)</code>.<br> Where <code>I</code> is the grayscale value of the RGB image: <br><code>I = 0.2989*R + 0.5870*G + 0.1140*B</code>.</td>
