AI algorithms for autonomous vehicles and the challenges they face

Self-driving cars are starting to become a reality in fields such as agriculture, transportation, and the military, and the day is fast approaching when ordinary consumers use them in their daily lives. The self-driving car performs the necessary actions based on sensor information and AI algorithms. It needs to collect data, plan trajectories, and execute driving routes. And these tasks, especially planning and executing trajectories, require unconventional programming methods that rely on machine learning techniques in AI.

Self-driving cars are starting to become a reality in fields such as agriculture, transportation, and the military, and the day is fast approaching when ordinary consumers use them in their daily lives. The self-driving car performs the necessary actions based on sensor information and AI algorithms. It needs to collect data, plan trajectories, and execute driving routes. And these tasks, especially planning and executing trajectories, require unconventional programming methods that rely on machine learning techniques in AI.

3. AI Algorithms Used in Self-Driving Vehicles

3.1. Path Planning and Control Algorithms

Traditional heuristic algorithms in computer science can be used for path planning and control, such as Bellman-Ford algorithm and Dijkstra algorithm. To apply these algorithms, the vehicle needs to be positioned throughout the entire process, and the positioning is done through sensors such as GPS, as well as simultaneous localization and mapping (SLAM) technology.

SLAM can be used in places where GPS signals cannot be searched, such as underground or enclosed spaces; it generates a map of the environment consisting of landmarks or obstacles, while estimating the state of the vehicle. SLAM is suitable for applications that cannot use maps but need to create maps. It uses sensors and special algorithms to create data models and generate maps.

3.2. Target detection algorithm

Object detection is one of the most important tasks that AI needs to handle in moving vehicles. Its algorithm has always been the direction of active research in the industry. Object detection relies on different sensors such as cameras or lidar, radar or other types of sensors. It usually uses deep learning algorithms and some type of neural network to get the job done.

Object detection has to be fast because it deals with a series of consecutive images generated by the vehicle as it moves.

Some of the latest techniques are based on Convolutional Neural Networks (CNN), such as R-CNN, Fast R-CNN, and You Only Look Once (Yolo) method. The R-CNN method first finds regions in the image that contain potential objects, and then tries to analyze each region, so it is somewhat slow, which is why the Fast R-CNN method and the Yolo method appear. The Yolo method uses a single convolutional neural network to find regions and simultaneously classify objects in the region, so it is much faster than other methods. Also, Yolo is able to see the entire image, which does not cause problems that R-CNN methods can encounter, such as mistaking the background image for the target object.

3.3. Decision Algorithms

Decision-making refers to determining the vehicle’s actions based on information from sensors. The vehicle continuously makes decisions based on its rules and environment, and the algorithms used for decision-making include:

decision tree

Support Vector Machine (SVM) Regression

deep reinforcement learning

4. Challenges for AI in autonomous vehicles

In addition to facing some of the same challenges as other AI applications, AI in the self-driving car space also faces some special hurdles involving concepts such as real-time, safety, and machine ethics. The following subsections describe each of these in detail.

4.1 Real-time response

A real-time system is a special kind of embedded system that is characterized by producing outputs or reactions within a certain time limit. They must be deterministic and minimalistic in design so that the expected real-time behavior is always met. To this end, they typically employ special real-time operating systems (RTOS) or bare-metal executors that interact directly with the hardware, avoiding the use of interpreted languages ​​and dynamic memory allocation. In some cases, real-time systems will deliberately employ a subset of programming languages ​​for speed and determinism.

AI solutions are often the opposite, using high-level programming languages, relying entirely on the operating system, employing techniques such as dynamic memory allocation and garbage collection. This creates challenges for the system to make real-time decisions.

In addition, whether the system is centralized or distributed also has an impact. Centralized systems are easier to build, but rely heavily on internal communications and a powerful central processor. Distributed systems use dedicated CPUs that can handle different subsystems and sensors, reducing the need for a complex central CPU. Moreover, the distributed architecture system consumes less power, and is also more flexible and less expensive.

4.2 Computational Complexity

Due to the large amount of data and the computational complexity, artificial intelligence algorithms (especially deep learning) require special hardware solutions. Such as deep optimization of hardware such as graphics processing unit (GPU) or tensor processing unit (TPU) to achieve fast parallel computing. However, speed usually comes at the expense of higher energy consumption and higher cost. Even with dedicated hardware, there is no certainty that a particular algorithm will solve a problem in real time. Therefore, given its complexity and CPU requirements, the choice of algorithm becomes an important factor affecting real-time systems. Measuring CPU compute and memory usage helps determine whether the algorithm will fit in a typical CPU in a modern vehicle.

4.3 Black Box Behavior

AI algorithms have been criticized for being harder to analyze than standard computer algorithms. This is because AI algorithms are more complex and rely on large amounts of data. A complex neural network can perform multiple AI tasks without needing to understand its control process.

In fact, deep neural networks can approximate essentially any function, which is why it is so successful. Deep neural networks can have tens of thousands of nodes that can be trained to reach a specific state. Also, they can have many hidden layers and many inputs and outputs. But when the system crashes, all of which can cause problems and forensic analysis is required to find out the exact cause of the crash. This is also an area of ​​active research in the industry and, if successful, will address many of the legal and technical issues associated with the use of AI.

A hybrid approach that combines AI with traditional control algorithms is one solution. This issue is especially important for accident prediction and handling. Why make this decision and not another? The entire decision-making process has to be demonstrated by AI algorithms, but it will be difficult or almost impossible to prove if the algorithm is only analyzed and viewed as a black box.

4.4 Accuracy and reliability

Computer vision for self-driving vehicles may not be ready yet, and while they can work in pristine conditions, they fail with even small disturbances on the sensor inputs. The training of AI algorithms is carried out in training data with certain characteristics, and its progress is slow.

Altering the data can cause classification and prediction algorithms to dramatically change their behavior, with catastrophic consequences. For example, a person carrying a large bag may not be perceived by the vehicle as a person.

Furthermore, it is difficult to predict what will be triggered by arbitrary data entering the system. This flaw has been exploited by attackers who use image data invisible to humans to trick deep learning algorithms into stopping neural networks from classifying correctly. Clearly, we need a lot of improvements to improve the reliability and accuracy of machine learning algorithms.

4.5 Security

The complexity of AI can also lead to security concerns. The more complex the system, the harder it is to develop, test, and deploy. But that’s only part of the problem. Another reality is that new safety standards are being introduced gradually but not yet adopted by the transportation industry.

To complicate matters further, the analysis and validation of AI systems is still in its early stages and its complexity is difficult to navigate. In general, formal verification is difficult even for traditional software systems, and even more difficult for AI software. Not just for autonomous vehicles, but for all AI applications, we must create robust, validated and validated AI solutions.

4.6 Security and AI

Artificial intelligence systems are critical to autonomous vehicles, and their safety is directly related to the robustness of the entire system. But machine learning systems are vulnerable, and there are already many such cases.

The biggest threat to AI is that hackers can manipulate the data that sensors send to the vehicle, allowing the system to make bad decisions. An experiment using a deep learning model to recognize road signs showed that “just adding a few black and white stickers to a stop sign can trick the algorithm into thinking it’s a 45 mph speed limit sign”. Training the system on data with certain adversarial features can improve this situation.

The biggest threat to the operation of AI in self-driving cars is interfering with the operation of sensors, which can cause changes in the sensor data flow to completely confuse the AI ​​algorithm. Solutions the researchers are exploring include hybrid systems, fuzzy logic, and traditional model-based control systems to complement AI. But the stakes are high because AI is already in some vehicle subsystems, and in every car in the near future.

4.7 Ethics and AI

Currently, the implementation of AI in autonomous and semi-autonomous vehicles is not yet ethically mature and has not been vigorously developed. Ethical values ​​are inherently human qualities that are difficult to formalize and implement in machines. In other words, deciding what is right and what is wrong is a vague concept for a machine. The full development of machine ethics may still be decades away. It is also an active area of ​​research that falls at the intersection between human psychology, machine learning, and social policy. As the technology matures, some regulations may emerge to help.

5 Conclusion

AI solutions are everywhere today. They’re part of Alexa and Google Home devices, part of lawn-mowing robots, and part of the self-driving vehicles we’re expected to use soon. The challenges and achievements of AI are the same in many industries, but the transportation industry, especially autonomous vehicles, faces special challenges in terms of safety and protection. This is partly due to the complexity of the task of creating an autonomous vehicle, and partly due to the state of the art in the field of AI and its suitability to solve complex problems such as object detection, path planning, and real-time decision-making. As more and more autonomous smart solutions penetrate more areas of the economy, many obstacles will be solved. Perhaps the right question we should be asking the industry is “Are we going to implement it safely?” and not just “Are we there?”

The Links:   PM100RSA060 BSM75GB170DN2

Author: Yoyokuo