From perception to decision-making, algorithms give autonomous driving a “backbone”

The second half of the trend towards widespread application of autonomous driving has already begun. At present, the national and local governments are continuously strengthening their investment in the construction of smart transportation demonstration zones such as intelligent networking, autonomous driving, and intelligent buses. At the same time, they are orderly promoting the formulation and implementation of various relevant standards, actively encouraging social enterprises to participate, and creating a good atmosphere for the application of relevant transportation technologies and the implementation of scenarios.

Recently, two major events have occurred in the autonomous driving industry, attracting the attention and discussion of industry professionals.

On December 7th, Uber announced the sale of its long losing autonomous driving business, marking the end of Uber's autonomous driving team ATG, which was established in 2015. The five-year development of autonomous driving has also come to an end.

On December 8th, Baidu held the Apollo Ecology Conference in Guangzhou and released the ANP (Apollo Navigation Pilot) for assisted driving. Apollo Lite is an urban road closed-loop solution based on L4 level pure visual autonomous driving, which will be launched in mass production models next year. ANP collaborates with the previously launched AVP parking plan, and Baidu plans to achieve front-end installation of 1 million vehicle models in the next 5 years.

The launch of a systematic solution for autonomous driving is good news for the majority of automotive users. In the process of implementing autonomous driving related applications, the role played by algorithms, cloud computing, etc. is becoming increasingly prominent. Today, machine learning algorithms are being used on a large scale to solve the increasing problems in the autonomous vehicle industry. Combining ECU (Electronic Control Unit) sensor data, people must strengthen the utilization of machine learning methods to meet new challenges. Potential applications equipped with machine learning algorithms include utilizing sensors distributed inside and outside the vehicle, such as radar, cameras, and laser detection. After integrating various types of data, it can be applied to sub scenarios such as driver driving condition assessment, driving route planning, and obstacle avoidance.

According to the research history of vehicle navigation systems, vehicle path planning algorithms are generally divided into static path planning algorithms and dynamic path planning algorithms. Static path planning, which seeks short paths based on physical geographic information and traffic rules as constraints, has become increasingly mature. Although it is relatively simple, its application significance is not significant for actual traffic conditions; Dynamic path planning is based on static path planning and combines real-time traffic information to adjust the pre planned optimal driving route in a timely manner until reaching the destination and obtaining the optimal path. The integration of algorithms into dynamic path planning has become a major branch of autonomous driving research.

In advanced driving assistance systems (ADAS), images obtained using sensors contain a variety of environmental data. At this point, filtering the image becomes very necessary, which can eliminate some irrelevant samples to obtain instance data for classification. Before classification, the key step is pattern recognition on a dataset, which is called a data reduction algorithm.

Due to changes in application scenarios, redefining the software architecture has led to changes in hardware architecture, with the core being the computation of artificial intelligence. In fact, from logical computing in the CPU era to intelligent computing deeply rooted in network computing, the architecture of hardware processors has been redefined, which is also applicable in smart cars and autonomous taxis.

Some experts pointed out that autonomous vehicle rely on the cooperation of radar, monitoring devices, visual computing, artificial intelligence and positioning systems, so that computers can automatically and safely operate motor vehicles without any human initiative. In the process of autonomous driving, cars need to achieve artificial intelligence through a series of processes such as perception, planning, control, and decision-making. In brief, the auto drive system can only make specific behavior decisions such as stopping, overtaking, lane changing, etc. according to the feedback of environmental information obtained from the perception fusion module, such as traffic lights, pedestrians, other vehicles and other data.

In theory, since autonomous vehicle are not as easily distracted or tired as human drivers, they actually have the potential to reduce road traffic accidents. However, when encountering new situations, humans are able to respond quickly based on knowledge categories and rational decision-making. At this stage, human beings are still unable to train autonomous vehicle to be familiar with all possible traffic conditions. In order to realize the "accident free trajectory" operation of autonomous vehicle, we need to break through the bottlenecks in algorithm, technology, supervision and other aspects.

With the improvement of algorithms and the maturity of various technologies such as the Internet of Things, cloud computing and big data, it is believed that the promotion and application of autonomous vehicle will run more smoothly.

Prev: Can AI use fortune telling to analyze facial expressions? Don’t pay the intelligence tax