AI algorithms in self-driving technology

Implementing autonomous driving is simply from perception and decision making to execution. Perception means collecting information from the vehicle itself and the outside through various sensors, and decision-making means that the computing unit of the vehicle analyzes the above-acquired information according to specific algorithms and makes decisions suitable for the current scenario, followed by execution.

And the whole process algorithm is extremely important. Autonomous driving is one of the important application scenarios of artificial intelligence technology. Its technology can not be achieved without the large-scale deployment of algorithms, including feature extraction from the perception link to neural network decision-making. All need to rely on algorithm improvements to improve the accuracy of obstacle detection and decision-making capabilities in complex scenarios.

AI algorithms are the most critical part of supporting autonomous driving technology, and mainstream autonomous driving companies are currently using machine learning and artificial intelligence algorithms to achieve this. Massive amounts of data are the basis for machine learning as well as AI algorithms. The constantly optimized algorithms can use the data obtained from the previously mentioned sensors, V2X facilities and high-precision map information, as well as the data collected on driving behaviour, driving experience, driving rules, cases, and the surrounding environment to identify and eventually plan routes and manipulate driving.

In terms of technical aspects, autonomous driving domain algorithms can be divided into perception algorithms, fusion algorithms, decision algorithms and execution algorithms. Perception algorithms convert sensor data into the machine language of the vehicle's location scene, including object detection, recognition and tracking, 3D environment modelling, and motion estimation of objects.

The core task of the fusion algorithm is to quantitatively unify the data acquired by different sensors with different dimensions, such as image-based or point cloud-based. As the requirement of L2+ autonomous driving for multi-sensor fusion accuracy increases, the fusion algorithm will gradually be forward-oriented (pre-fusion), and its layers will gradually move forward from back-end components such as domain controllers to the sensor level. Fusion will be completed within the sensor to improve the efficiency of data processing.

Decision algorithms, that is, are based on the output results of the perception algorithm and give the final behavioural action instructions, including behavioural decisions such as car following, stopping and chasing, as well as action decisions such as car steering, speed, and path planning, etc.

Autonomous driving is divided into L0-L5 levels according to the degree of automation function. L1-L3 mainly plays the function of assisted driving. After the L4 level, vehicle control can be given to the artificial intelligence system.

Different levels require different functions and different algorithms. For example, L1's ACC adaptive cruise control, LKA lane departure assistance, AEB automatic braking, and BSM blind spot monitoring require ACC system control algorithms, LDW lane departure warning algorithms, LKA lane keeping assistance algorithms, AEB automatic braking algorithms, and BSM blind spot monitoring algorithms.

L3+, for example, requires a TJP traffic jam assist algorithm, HWP highway assist algorithm, city road autopilot algorithm, highway autopilot algorithm, AVP automatic parking algorithm, and L5 requires various types of autopilot algorithms and so on to achieve the corresponding functions.

The ability of different manufacturers to provide algorithms varies, such as traditional Tier1 manufacturers, Bosch, Continental, Dexiawei, and some software algorithm manufacturers, etc., can provide some single-function module algorithms, which can be better applied to L1-L2 assisted driving; then, for example, Momenta, Mini eye and other algorithm solution providers can provide a complete ADAS or autonomous driving solutions.