Search

To ensure the safe operation of highway traffic lines, given the imperfect feature extraction of existing road pit defect detection models and the practicability of detection equipment, this paper proposes a lightweight target detection algorithm with enhanced feature extraction based on the YOLO (You Only Look Once) algorithm. The BIFPN (Bidirectional Feature Pyramid Network) network structure is used for multi-scale feature fusion to enhance the feature extraction ability, and Varifocal Loss is used to optimize the sample imbalance problem, which improves the accuracy of road defect target detection. In the evaluation test of the model in the constructed PCD1 (Pavement Check Dataset) dataset, the mAP@.5 (mean Average Precision when IoU = 0.5) of the BV-YOLOv5S (BiFPN Varifocal Loss-YOLOv5S) model increased by 4.1%, 3%, and 0.9%, respectively, compared with the YOLOv3-tiny, YOLOv5S, and B-YOLOv5S (BiFPN-YOLOv5S; BV-YOLOv5S does not use the Improved Focal Loss function) models. Through the analysis and comparison of experimental results, it is proved that the proposed BV-YOLOv5S network model performs better and is more reliable in the detection of pavement defects and can meet the needs of road safety detection projects with high real-time and flexibility requirements. [ABSTRACT FROM AUTHOR]

This paper proposes a real-time feature extraction VLSI architecture for high-resolution images based on the accelerated KAZE algorithm. Firstly, a new system architecture is proposed. It increases the system throughput, provides flexibility in image resolution, and offers trade-offs between speed and scaling robustness. The architecture consists of a two-dimensional pipeline array that fully utilizes computational similarities in octaves. Secondly, a substructure (block-serial discrete-time cellular neural network) that can realize a nonlinear filter is proposed. This structure decreases the memory demand through the removal of data dependency. Thirdly, a hardware-friendly descriptor is introduced in order to overcome the hardware design bottleneck through the polar sample pattern; a simplified method to realize rotation invariance is also presented. Finally, the proposed architecture is designed in TSMC 65 nm CMOS technology. The experimental results show a performance of 127 fps in full HD resolution at 200 MHz frequency. The peak performance reaches 181 GOPS and the throughput is double the speed of other state-of-the-art architectures. [ABSTRACT FROM AUTHOR]

Traffic sign detection systems provide important road control information for unmanned driving systems or auxiliary driving. In this paper, the Faster region with a convolutional neural network (R-CNN) for traffic sign detection in real traffic situations has been systematically improved. First, a first step region proposal algorithm based on simplified Gabor wavelets (SGWs) and maximally stable extremal regions (MSERs) is proposed. In this way, the region proposal a priori information is obtained and will be used for improving the Faster R-CNN. This part of our method is named as the highly possible regions proposal network (HP-RPN). Second, in order to solve the problem that the Faster R-CNN cannot effectively detect small targets, a method that combines the features of the third, fourth, and fifth layers of VGG16 to enrich the features of small targets is proposed. Third, the secondary region of interest method to enhance the feature of detection objects and improve the classification capability of the Faster R-CNN is proposed. Finally, a method of merging the German traffic sign detection benchmark (GTSDB) and Chinese traffic sign dataset (CTSD) databases into one larger database to increase the number of database samples is proposed. Experimental results show that our method improves the detection performance, especially for small targets. [ABSTRACT FROM AUTHOR]

A binocular vision system is a common perception component of an intelligent vehicle. Benefiting from the biomimetic structure, the system is simple and effective. Which are extremely snesitive on external factors, especially missing vision signals. In this paper, a virtual view-generation algorithm based on generative adversarial networks (GAN) is proposed to enhance the robustness of binocular vision systems. The proposed model consists of two parts: generative network and discriminator network. To improve the quality of a virtual view, a generative network structure based on 3D convolutional neural networks (3D-CNN) and attentive mechanisms is introduced to extract the time-series features from image sequences. To avoid gradient vanish during training, the dense block structure is utilized to improve the discriminator network. Meanwhile, three kinds of image features, including image edge, depth map and optical flow are extracted to constrain the supervised training of model. The final results on KITTI and Cityscapes datasets demonstrate that our algorithm outperforms conventional methods, and the missing vision signal can be replaced by a generated virtual view. [ABSTRACT FROM AUTHOR]

It is crucial for robots to autonomously steer in complex environments safely without colliding with any obstacles. Compared to conventional methods, deep reinforcement learning-based methods are able to learn from past experiences automatically and enhance the generalization capability to cope with unseen circumstances. Therefore, we propose an end-to-end deep reinforcement learning algorithm in this paper to improve the performance of autonomous steering in complex environments. By embedding a branching noisy dueling architecture, the proposed model is capable of deriving steering commands directly from raw depth images with high efficiency. Specifically, our learning-based approach extracts the feature representation from depth inputs through convolutional neural networks and maps it to both linear and angular velocity commands simultaneously through different streams of the network. Moreover, the training framework is also meticulously designed to improve the learning efficiency and effectiveness. It is worth noting that the developed system is readily transferable from virtual training scenarios to real-world deployment without any fine-tuning by utilizing depth images. The proposed method is evaluated and compared with a series of baseline methods in various virtual environments. Experimental results demonstrate the superiority of the proposed model in terms of average reward, learning efficiency, success rate as well as computational time. Moreover, a variety of real-world experiments are also conducted which reveal the high adaptability of our model to both static and dynamic obstacle-cluttered environments. [ABSTRACT FROM AUTHOR]