A brief overview of some noteworthy exploited Machine Learning techniques to address the self-driving challenges.
As mentioned in the introduction article, at present, a number of Machine Learning (ML) techniques is already implemented in the ADAS systems. For instance, in some cars, facial recognition is deployed to monitor and notify when the driver does not pay attention to the road. In a near future, such a technology may also reasonably be used to prevent thefts by allowing the control of the car exclusively to authorised parties.
However, the concept of Autonomous Vehicles (AVs), on which the blog develops, is based on a deeper automation level wherein drivers are entirely replaced.
The challenges to address for achieving self-driving can be categorised into 4 main key components, namely Perception, Localisation, Planning and Control.
Perception
An autonomous system must be able to extract and collect critical information to develop a comprehension of the surrounding environment. The devices employed in AVs to retrieve data are usually cameras, LiDAR and radar sensors.
Unlike the others, cameras sense colours, detect small targets and are passive, i.e. the measurements are performed without emitting signals. As they replicate human vision, cameras are particularly useful to detect and classify traffic lights, lane lines, road signs and objects.
Overall, Deep Learning (DL) methods based on Convolutional Neural Networks (CNNs), e.g. LaneNet for lane line detection, are preferred to traditional ML algorithms since more efficient, accurate and faster. A popular approach for classifying the form and type of lane lines consists in segmenting the input image using a series of CNN layers that downsample and upsample it, such as in Variational Autoencoders (VAEs) or U-Net architectures.
Object detection is paramount for identifying the location, size and class of the entities populating the road environment. However, CNNs can usually classify objects that occupy a sizable portion of the image. To solve this problem avoiding unnecessary computationally expensive sliding windows, YOLO and SSD state-of-the-art models directly receive an entire image divided into a grid as input. The network then returns a class probability for each grid cell.
LiDAR and radar active sensors create a 3D point cloud map of the environment by computing the elapsed time between the transmission and reception of a specific signal. Whereas the radar emits radio waves, the LiDAR sensor sends out laser pulses. Differently to cameras, both of them are not affected by the illumination of the environment. As radio waves are less absorbed than light waves by the objects that bounce them back, radars are particularly suited to work on long-range distances. Nonetheless, they notoriously remain less accurate than LiDAR sensors.
Recently, there is evidence of an increasing interest of researchers in 3D object detection using LiDAR. Some examples are VoxelNet, PointPillars and PIXOR. However, such deep learning studies are still in their early stages.
The best environmental perception performance in the literature is achieved by increasing the method robustness through the combination of distinct sensor typologies.
Localisation
In this blog, localisation is slightly neglected since less attractive than the other elements. In few words, it involves determining the position and orientation of the vehicle with respect to the environment in which it moves. Specifically, the GPS alone is not enough satisfactory because it is inaccurate and can be influenced by weather conditions.
References
Cohen, Jeremy. "Deep Learning in Self-Driving Cars - Becoming Human: Artificial Intelligence Magazine." Medium, 12 Mar. 2021, becominghuman.ai/deep-learning-algorithms-in-self-driving-cars-14b13a895068.
Cohen, Jeremy. "Computer Vision Applications in Self-Driving Cars - Becoming Human: Artificial Intelligence Magazine." Medium, 14 Oct. 2020, becominghuman.ai/computer-vision-applications-in-self-driving-cars-610561e14118.
Osman Ors, Ali. "The Role of Machine Learning in Autonomous Vehicles." www.electronicdesign.com, 3 Dec. 2020.
Swarit Dholakia. "Perception: How Self-Driving Cars ‘See’ the World." Medium, 14 Mar. 2020, swarit.medium.com/perception-how-self-driving-cars-see-the-world-ae630636f4c.
Cohen, Jeremy. "Sensor Fusion." Medium, 2021, towardsdatascience.com/sensor-fusion-90135614fde6.
Yurtsever, Ekim, et al. "A survey of autonomous driving: Common practices and emerging technologies." IEEE Access 8 (2020): 58443-58469.
Pendleton, Scott Drew, et al. "Perception, planning, control, and coordination for autonomous vehicles." Machines 5.1 (2017): 6.
Lai, Albert. "How Do Self-Driving Cars See? - Towards Data Science." Medium, 15 Dec. 2018, towardsdatascience.com/how-do-self-driving-cars-see-13054aee2503.
Wang, Ze, Weiqiang Ren, and Qiang Qiu. "Lanenet: Real-time lane detection networks for autonomous driving." arXiv preprint arXiv:1807.01726 (2018).
Zhou, Yin, and Oncel Tuzel. "Voxelnet: End-to-end learning for point cloud based 3d object detection." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018.
Yang, Bin, Wenjie Luo, and Raquel Urtasun. "Pixor: Real-time 3d object detection from point clouds." Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. 2018.
Lang, Alex H., et al. "Pointpillars: Fast encoders for object detection from point clouds." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019.
The images in the blog are either copyright free or designed from scratch. The gif regarding the YOLO model derives from a video published on their official website:
Comentarios