Heavy Vehicle Detection Using Fine-Tuned Deep Learning
16 May 2018-pp 1903-1910
TL;DR: Experiments show that vehicle’s make and model can be recognized from transportation images effectively by using the proposed fine-tuned detection system, and demonstrate that the proposed detection system performs accurately with other simple and complex scenarios in detecting heavy vehicles in comparison with past vehicle detection systems.
Abstract: Heavy vehicles develop technical snag and traffic jam on streets. Accidents between heavy vehicle and road users, for example, pedestrians often result in severe injuries of the weaker street users. The highway safety and traffic jams can be secured with detection of heavy and overloaded vehicles on the highway to facilitate light motor vehicles like cars, scooters. A model for heavy vehicle detection using fine-tuned based on deep learning is proposed to deal with entangled transportation scene. This model comprises two parts, vehicle detection model and vehicle fine-grained detection. This step provides data for the next classification model. Experiments show that vehicle’s make and model can be recognized from transportation images effectively by using our method. Experimental results demonstrate that the proposed detection system performs accurately with other simple and complex scenarios in detecting heavy vehicles in comparison with past vehicle detection systems.
References
More filters
25 Oct 2012
TL;DR: A novel, lightweight approach to real-time detection of vehicles using parts at intersections using strong classifiers trained with active learning and a comparison of detection results using geometric image features and appearance-based features.
Abstract: In this study, we propose a novel, lightweight approach to real-time detection of vehicles using parts at intersections. Intersections feature oncoming, preceding, and cross traffic, which presents challenges for vision-based vehicle detection. Ubiquitous partial occlusions further complicate the vehicle detection task, and occur when vehicles enter and leave the camera's field of view. To confront these issues, we independently detect vehicle parts using strong classifiers trained with active learning. We match part responses using a learned matching classification. The learning process for part configurations leverages user input regarding full vehicle configurations. Part configurations are evaluated using Support Vector Machine classification. We present a comparison of detection results using geometric image features and appearance-based features. The full vehicle detection by parts has been evaluated on real-world data, runs in real time, and shows promise for future work in urban driver assistance.
51 citations
01 May 2016
TL;DR: A new approach to detect traffic signs based on cascaded convolutional neural networks (CNNs), where the local binary pattern (LBP) feature detector and the AdaBoost classifier are combined to extract regions of interest (ROI) for coarse selection.
Abstract: In this paper, we present a new approach to detect traffic signs based on cascaded convolutional neural networks (CNNs). First, the local binary pattern (LBP) feature detector and the AdaBoost classifier are combined to extract regions of interest (ROI) for coarse selection. Next, cascaded CNNs are employed to reduce negative samples of ROI for traffic sign recognition. Compared with the conventional CNN, our CNN contains three convolutional layers and its classification part is replaced by the support vector machine (SVM). The German traffic sign detection benchmark is used and experimental results demonstrate that the proposed method can achieve competitive results when compared with the state-of-the-art approaches.
31 citations
TL;DR: The fire localization is taken as a nonlinear bearing-only tracking issue for the case where the covariance of measurement noise is unknown and a specific variational Bayesian adaptive square-cubature Kalman filter is proposed to estimate the coordinate of the center.
Abstract: Fire localization problem is studied based on temperature data taken by wireless sensor arrays and a novel range-range-range (RRR) model is proposed to overcome shortcomings, which exists in the current range-point-range (RPR) model in this paper. For a single sensor array composed of four sensors deployed with a square, three angle estimates on fire bearing can be obtained using far-field localization technology. These angle estimates are used to get their statistical mean and variance at a single time. Based on the statistical features, we propose two fire localization methods under the RRR frame, which are angle bisector and nonlinear filtering methods. For the angle bisector method, a recursive formula of the mean and variance is presented in time series so that global angle estimates can be used. Furthermore, a fire coordinate estimate, which is actually the center of estimated-range circle, can be taken by use of intersecting two angle bisectors from two sensor arrays. Moreover, the estimation of a radius for the estimated fire region is also realized. In order to improve localization accuracy and robustness of fire estimation to non-Gaussian noise component, the fire localization is taken as a nonlinear bearing-only tracking issue for the case where the covariance of measurement noise is unknown and a specific variational Bayesian adaptive square-cubature Kalman filter is proposed to estimate the coordinate of the center. These proposed algorithms not only provide some new points of view on the fire localization for limited interior space, but are helpful for practical fire fighting applications.
19 citations