Towards Efficient Risky Driving Detection: A Benchmark and a Semi-Supervised Model

Risky driving is a major factor in traffic incidents, necessitating constant monitoring and prevention through Intelligent Transportation Systems (ITS). Despite recent progress, a lack of suitable data for detecting risky driving in traffic surveillance settings remains a significant challenge. To address this issue, Bayonet-Drivers, a pioneering benchmark for risky driving detection, is proposed. The unique challenge posed by Bayonet-Drivers arises from the nature of the original data obtained from intelligent monitoring and recording systems, rather than in-vehicle cameras. Bayonet-Drivers encompasses a broad spectrum of challenging scenarios, thereby enhancing the resilience and generalizability of algorithms for detecting risky driving. Further, to address the scarcity of labeled data without compromising detection accuracy, a novel semi-supervised network architecture, named DGMB-Net, is proposed. Within DGMB-Net, an enhanced semi-supervised method founded on a teacher-student model is introduced, aiming at bypassing the time-consuming and labor-intensive tasks associated with data labeling. Additionally, DGMB-Net has engineered an Adaptive Perceptual Learning (APL) Module and a Hierarchical Feature Pyramid Network (HFPN) to amplify spatial perception capabilities and amalgamate features at varying scales and levels, thus boosting detection precision. Extensive experiments on widely utilized datasets, including the State Farm dataset and Bayonet-Drivers, demonstrated the remarkable performance of the proposed DGMB-Net.

The Braking-Pressure and Driving-Direction Determination System (BDDS) Using Road Roughness and Passenger Conditions of Surrounding Vehicles

A fully autonomous vehicle must ensure not only fully autonomous driving but also the safety and comfort of its passengers. However, the self-driving technology that is currently completed focuses only on perfect driving and does not guarantee the safety and comfort of passengers. This paper proposes a braking-pressure and driving-direction determination system (BDDS), which computes the brake pressure and steering angle optimized for passenger safety by utilizing more diverse information than existing autonomous vehicles. The BDDS proposed in this paper consists of two modules. The road roughness classification module (RRCM) classifies the roughness of the road by using the pressure data applied to the suspension and the K-NN algorithm and computes the optimal brake pressure. The passenger recognition and sharing module (PRSM) identifies the current occupant status of the vehicle by using a body pressure sensor and CNN, shares the information with surrounding vehicles, and computes the optimal steering angle using passenger information and road information. As a result of the simulations described in this paper, the parameters of AI models were optimized. In addition, the RRCS was about 7% more accurate than the K-means clustering algorithm, and PRS was about 9% more accurate than the existing seat recognition system.

The difficulty of computing stable and accurate neural networks: On the barriers of deep learning and Smale's 18th problem

Instability is the Achilles’ heel of modern artificial intelligence (AI) and a paradox, with training algorithms finding unstable neural networks (NNs) despite the existence of stable ones. This foundational issue relates to Smale’s 18th mathematical problem for the 21st century on the limits of AI. By expanding methodologies initiated by Gödel and Turing, we demonstrate limitations on the existence of (even randomized) algorithms for computing NNs. Despite numerous existence results of NNs with great approximation properties, only in specific cases do there also exist algorithms that can compute them. We initiate a classification theory on which NNs can be trained and introduce NNs that—under suitable conditions—are robust to perturbations and exponentially accurate in the number of hidden layers.

Deep learning (DL) has had unprecedented success and is now entering scientific computing with full force. However, current DL methods typically suffer from instability, even when universal approximation properties guarantee the existence of stable neural networks (NNs). We address this paradox by demonstrating basic well-conditioned problems in scientific computing where one can prove the existence of NNs with great approximation qualities; however, there does not exist any algorithm, even randomized, that can train (or compute) such a NN. For any positive integers K>2 and L, there are cases where simultaneously 1) no randomized training algorithm can compute a NN correct to K digits with probability greater than 1/2; 2) there exists a deterministic training algorithm that computes a NN with K –1 correct digits, but any such (even randomized) algorithm needs arbitrarily many training data; and 3) there exists a deterministic training algorithm that computes a NN with K –2 correct digits using no more than L training samples. These results imply a classification theory describing conditions under which (stable) NNs with a given accuracy can be computed by an algorithm. We begin this theory by establishing sufficient conditions for the existence of algorithms that compute stable NNs in inverse problems. We introduce fast iterative restarted networks (FIRENETs), which we both prove and numerically verify are stable. Moreover, we prove that only O(|log (ϵ)|) layers are needed for an ϵ-accurate solution to the inverse problem.