Object detection in autonomous driving
Autonomous cars have to perceive environment, predict behavior of environment items, make decisions and planning and bring it to life. They operate in uncontrolled or complex world. So the perception is essential stage, resulting in giving objects, which are aimed at being processed in the next stage. Perception systems must give precise information, be robust in different outdoor conditions or when some part break down and be real time.
The purpose of object detection in autonomous driving is to provide the possibility to drive and achieve the target, avoid collision with obstacles and don’t provoke other traffic participants to change their intention (directory or speed). So object detection assumes the necessity to know the objects positions.
The object detection flow can be divided in the following steps:
- Perception. First of all functionality of sensor drivers, which provide data representation and interaction with operating systems, should be covered. Also phase of preprocessing should be taken into account. The aim of this phase is to minimize data, but keeping them understandable. Preprocessing basically removes data, which are not useful for further steps. For example useless data are those points, which are out of distance or angle range, as they have no context and interfere with another layer respectively. Besides it, preprocessing filters out redundant and dirty data. An example of dealing with redundant data is a reducing point cloud to representative set of points. Lastly, preprocessing makes data consistent. It means namely data fusion from different source and putting all time aligned messages into single coordinate system. Since cars have multi-modal sensors with various sensing modalities, it is necessary to fuse their data.
- Processing/Understanding. First step is ground filtering, as ground points are not essential for avoiding hitting of objects (object detection). This approach improves robustness of clustering, which is the next step. The idea of clustering is divide point cloud into separate objects. Point cloud, being big homogeneous structure, can’t be used itself to find objects to avoid hitting them, as it is impossible to track individual objects or intersection based planning. To pick out objects segmentation and clustering are applied. The last step is shape extraction. We need it as point blobs, which were extracted during the previous step of segmentation and clustering, cause extra memory and computational load, when the typical task such as tracking is performed. As shape can not be found unambiguously, its criteria should be easiness of collision check, computing and rationality of actual object representation.
- Usage of detected object. One of the usage of detected objects is tracking. In order to have predictions how the extracted object is going to behave, it should be classified. Another usage is collision detection. Also extracted object can be combined with data from other sensors or used as region of interest.
Below are listed the sensors, which are used in autonomous driving:
- Radar measures the radio wave signal runtime, detecting objects position. Also it can evaluate the object’s velocity by the Doppler effect. Advantage is robustness to lighting and weather conditions. Disadvantage is low resolution, which leads to difficulty of object classification. Adaptive cruise control (ACC) and traffic jam assistance systems are applications for Radars.
- IMU (Inertial Measurement Units) and Odometers measure cars internal information, vehicles accelerations and odometery respectively. Their application is vehicle dynamic driving control and accurate localization.
- Visual and thermal cameras are good for getting texture information of a car surroundings. For visual cameras lighting and weather conditions are essential, whereas thermal cameras are not so sensitive to day and night changes, due to their possibility to detect infrared radiation, produced by object heat.
- GNSS provides precise localization for autonomous vehicles by a global satellite system and together with HD Maps is used path planning.
- Lidar provide precise depth information as 3D points, by measuring time of reflection laser beams, which are regularly emitted. The disadvantages is difficulty of getting information on textures (hence object classification) and of sparseness of points of remote objects. The advantage of Lidar is robustness to lightening and weather conditions.
- Ultrasonic sensors emit high-frequency sound waves for measuring distance to objects. The field of application is near-range object detection in low speed scenarios (parking). Disadvantage is low robustness to environment conditions.