Real-time running deep convolutional neural networks on embedded electronics is one recent focus for distracted driving detection. In this work, we proposed $\textit{FRNet}$ , which is a unique, efficient, and real-time architecture. The $\textit{FRNet}$ converts spatially distributed features into depth distribution by a feature reorganization block. This block compresses the volume of backbone, reduces the memory read/write volumes along with multiply-accumulate operations, and extracts key features faster. In addition, a ultra-lightweight backbone was designed, with an atypical reshape strategy. This atypical strategy designed based on pixel-level analysis, for compensating the accuracy decline along with feature reorganization. The proposed $\textit{FRNet}$  offered excellent real-time performance on low power embedded platforms, and has a competitive accuracy with previous state-of-the-art models. It achieved 97.55% accuracy on the $\textit{SFD+AUCDD-V1}$ , and 99.86% on $\textit{3MDAD}$ . On an automotive-grade embedded demo board, it costs 16.93 ms per frame and achieved 59 FPS. As of today, this is the fastest record for end-to-end distraction detection. Experiments revealed the real-time and accuracy are best balanced with FRNet . The model is publicly available at https://github.com/congduan-HNU/FRNet.

Road garbage recognition and cleanliness assessment are very important for intelligent cleaning of vehicles. However, widely used garbage detection methods cannot provide accurate information to assess cleanliness. Also, only a few contributions are available in the road cleanliness assessment, and they too mainly depend on the category and quantity of garbage. In this paper, deep supervision UNet++ (DUNet++) is proposed to solve the problem of road garbage classification and segmentation, which can directly impact the output of the category of garbage and the occupied ground area. A new, simple and accurate method is established to assess road cleanliness by combining the results of garbage segmentation. Compared with the cleanliness index (CI), the road cleanliness index based on semantic segmentation (CISS) not only considers the road area occupied by different types or categories of the garbage but also considers the difficulty of cleaning them. A road garbage segmentation dataset, which consists of four categories (stones, leaves, sand, and bottles), is collected to train the designed garbage segmentation model, and it has achieved considerable garbage segmentation improvement with an MIoU (Mean Intersection over Union) of $76.73\pm 0.11$ . Especially when compared with the state-of-the-art methods, the model we designed has greatly improved the accuracy of garbage segmentation. Experimental results have shown that our road garbage segmentation and cleanliness assessment methods are relatively simple, and can provide accurate road cleanliness information for cleaning vehicles.

An intelligent cleaning vehicle improves the efficiency of road cleaning to a great extent. In this case, road garbage recognition is fundamental and crucial. Usually, the pebbles are too small, and the sand has an inconspicuous boundary and unfixed shape on the roads. These issues make the stones and sand too hard to be detected by a garbage recognition system. Hence, we designed a novel semantic segmentation network that acquires the areas and categories of road garbage. We designed a Deep Supervision and High Fusion (DSHF) block to improve the road garbage segmentation accuracy. The designed block with a backbone network of HFCN and UNet++ is comparable with a model that only has either a deep supervision block or a high fusion block. According to the collected road garbage segmentation dataset comprising four categories (stones, leaves, sand and bottles), the model we designed has improved the metric value of MPA by 3% over the state-of-art methods and achieved the highest MIoU (Mean Intersection over Union) with the value of 77.92%. The results of our experiments show that the semantic segmentation of road garbage is feasible and that the proposed DSHF block is practical for improving the segmentation effect of road garbage.

Highways, as a type of high-grade road, are an essential component of intelligent connected vehicle testing and a crucial aspect in achieving fully autonomous driving technology. To effectively predict the driver behaviors when encountering merging vehicles in high-speed merging areas, this paper proposes a high-speed main lane vehicle driving behavior model that combines Convolutional Neural Network (CNN) based on Residual Structure with a Transformer network. This model takes input in the form of the target vehicle's motion state information and surrounding vehicle interaction data, and predicts the current driving behavior state of the target vehicle as well as the next-stage driving behavior. Finally, the effectiveness of this model is validated on the Next-Generation Simulation (NGSIM) dataset and the Exits and Entries Drone (exiD) dataset. Comparative experiments with the Time series Transformer (TST) model and the Multi-head Attention CNN-LSTM (MCNNLSTM) model are conducted. The results indicate that the proposed model outperforms other models in aspects such as driving behavior recognition, with a correct identification rate of 94% and 95% in the two major datasets, respectively. The driving behavior prediction model presented in this paper can assist intelligent connected vehicles in high-speed ramp merging decision-making and planning.

Left-behind humans inside the car or bus have caused a lot of accidents, so it is essential to detect the humans in vehicle. Current human detection methods rely on wearable devices, oxygen sensors, and special seat designs in vehicles, but those sensors cannot adapt to ever-changing environments. To solve those problems and especially to improve passengers’ safety on the bus, we propose a method to accomplishing human detection by fusion vision and microwave radar information in various environments in vehicle. For vision information, we use different networks to extract human and human face features, and fusion of the detection results in different models to improve human detection accuracy. The human detection model is MobileNet-V2, and the human face detection model is MTCNN. A new matching schedule and tracking objects management rule based on the Kernelized Correlation Filter tracker are designed to track the human and human face detection boxes. The microwave radar information is used to detect moving objects. Finally, the fusion vision and microwave radar detection results are implemented. Experiments show that our method has improved the human detection accuracy in vehicle, and this method can be used for detection of left-behind children on the school bus.

The around view monitor system (AVM) has been widely used in passenger vehicles. However, for articulated engineering vehicles (AEV), AVM is difficult to directly apply due to the long body and the articulation angle. In this paper, a flexible technique for variable-angle around view monitor system (VAVM) towards the AEV is proposed, which includes static calibration and dynamic execution. In the static calibration stage, the camera poses in the vehicle coordinate system with the joint of AEV as the coordinate origin can be calibrated using irregularly placed multiple checkerboards without knowing their relative position. This flexible calibration approach makes the variable-angle mapping from the fisheye image to the bird’s eye view image can be directly obtained by updating the camera poses on the tractor according to the articulation angle when the AEV turns. In the dynamic execution stage, an optimization algorithm based on minimizing the photometric error is designed to reduce the cumulative error of the angle sensor. Finally, all the bird’s eye view images are mosaiced and fused to a variable-angle around view image (VAVI). The experimental results show that the VAVM can output globally balanced and seamless VAVIs under different angles and visual environment conditions. Furthermore, the practicability of VAVM is also demonstrated by a real AEV.