Automatic detection and recognition of traffic signs plays a crucial role in management of the traffic-sign inventory. It provides an accurate and timely way to manage traffic-sign inventory with a minimal human effort. In the computer vision community, the recognition and detection of traffic signs are a well-researched problem. A vast majority of existing approaches perform well on traffic signs needed for advanced driver-assistance and autonomous systems. However, this represents a relatively small number of all traffic signs (around 50 categories out of several hundred) and performance on the remaining set of traffic signs, which are required to eliminate the manual labor in traffic-sign inventory management, remains an open question. In this paper, we address the issue of detecting and recognizing a large number of traffic-sign categories suitable for automating traffic-sign inventory management. We adopt a convolutional neural network (CNN) approach, the mask R-CNN, to address the full pipeline of detection and recognition with automatic end-to-end learning. We propose several improvements that are evaluated on the detection of traffic signs and result in an improved overall performance. This approach is applied to detection of 200 traffic-sign categories represented in our novel dataset. The results are reported on highly challenging traffic-sign categories that have not yet been considered in previous works. We provide comprehensive analysis of the deep learning method for the detection of traffic signs with a large intra-category appearance variation and show below 3% error rates with the proposed approach, which is sufficient for deployment in practical applications of the traffic-sign inventory management.

Robust local navigation is a critical capability for any mobile robot operating in a real-world, unstructured environment, especially when there are humans or other moving obstacles in the workspace. One of the most commonly used methods for local navigation is the Dynamic Window Approach (DWA), which does not address the problem of dynamic obstacles and depends heavily on the settings of the parameters in its cost function. Thus, it is a static approach that does not adapt to the characteristics of the environment, which can change significantly. On the other hand, data-driven deep learning approaches attempt to adapt to the characteristics of the environment by predicting the appropriate robot motion based on the current observation. However, they cannot guarantee collision-free trajectories for unseen inputs. In this work, we combine the best of both worlds. We propose a neural network to predict the weights of the DWA, which is then used for safe local navigation. To address the problem of dynamic obstacles the proposed method considers a short sequence of observations to allow the network to model the motion of the obstacles and adjust the DWA weights accordingly. The network is trained using the Proximal Policy Optimization (PPO) in a reinforcement learning setting in a simulated dynamic environment. We perform a comprehensive evaluation of the proposed approach in realistic scenarios using range scans of real 3D spaces and show that it outperforms both DWA and purely Deep Learning approaches.

Recent trends in the development of autonomous vehicles focus on real-time processing of vast amounts of data from various sensors. The data can be acquired using multiple cameras, lidars, ultrasonic sensors, and radars to collect useful information about the state of the traffic and the surroundings. Significant computational power is required to process the data fast enough, and this is even more pronounced in vehicles that not only assist the driver but are capable of fully autonomous driving. This article proposes speed and accuracy improvement of traffic sign detection and recognition in high-definition images, based on focusing on different regions of interest in traffic images. These regions are determined with efficient and parallelized preprocessing of every traffic image, after which convolutional neural network is applied for detection and recognition in parallel on graphics processing units. We employed different “You Only Look Once” (YOLO) architectures as baseline detectors, due to their speed, straightforward architecture, and high accuracy in general object detection tasks. Several preprocessing procedures were proposed, to achieve real-time performance requirement. Our experiments using a large-scale traffic sign dataset show that we can achieve real-time detection in high-definition images with high recognition accuracy.