29 July 2020 in IEEE Transactions on Intelligent Transportation Systems
Deep learning has achieved good performance in short-term traffic forecasting recently. However, the stochasticity and distribution imbalance are main characteristics to traffic flow, and these will bring the uncertainty and induce the network overfitting problem during deep learning. To deal with the problems, a new end-to-end hybrid deep learning network model, named M-B-LSTM, is proposed for short-term traffic flow forecasting in this paper. In the M-B-LSTM model, an online self-learning network is constructed as a data mapping layer to learn and equalize the traffic flow statistic distribution for reducing the effect of distribution imbalance and overfitting problem during network learning. Besides, the deep bidirectional long short-term memory network (DBLSTM) is introduced to reduce the uncertainty problem by forward and reverse contexts approximation process in the stochasticity reducing layer, and then the long short-term memory network (LSTM) is used to forecast the next traffic flow state in the forecasting layer. Furthermore, sufficient comparative experiments have been conducted and the results show the proposed model has better ability on solving uncertainty and overfitting problems than the state-of-art methods.
15 April 2021 in IEEE Transactions on Vehicular Technology
Autonomous vehicles have to share the road with human-driven ones for a prolonged period in the near future. Thus, they should have the ability to understand the lane changing intentions of surrounding human-driven vehicles. Unlike human drivers, the state-of-the-art models used in autonomous vehicles recognize a driver's lane changing intention based on the target vehicle's lateral movement, which leaves little to no time for the autonomous vehicle to react. In this paper, a Human-like Lane Changing Intention Understanding Model (HLCIUM) for autonomous driving is proposed to understand the lane changing intentions of surrounding vehicles. By imitating the selective attention mechanism of human vision systems, the proposed model emulates the way human drivers concentrate on the surrounding vehicles and recognizes their lane changing intentions accordingly. The velocity changes of the surrounding vehicles are treated as the lane changing hints, and the attention is drawn to the corresponding vehicle following a saliency-based scheme. Then, the lane changing intention is identified by a Hidden Markov Model (HMM) based intention recognizer. The proposed model is tested with Next Generation Simulation (NGSIM) vehicle trajectory dataset, and the proposed method has reached 90.89% in detecting lane changing intention and 88.58% in lane keeping in urban road scenarios, and reached 87.73% and 87.48% for the lane changing and lane keeping intentions in highway scenarios, respectively. Importantly, the average recognition time before the lane changing maneuver of the proposed model is 6.67 seconds for the urban road datasets and 7.08 seconds for the highway datasets, which is far earlier than state-of-the-art models. Furthermore, the proposed method shows efficiency and robustness in complex real urban traffic datasets, which is ideal to use in human-like autonomous driving systems.
16 September 2018 in IEEE Access
Traffic signal control is widely used at intersection to improve its operation efficiency. However, the existing signal control systems cannot satisfy the control requirements under unsaturated, saturated, and oversaturated conditions, which will induce queue spillover, even network deadlock. A signal Cooperative Control method with traffic Supply and Demand (CCSD) on a single intersection is put forward to maximize the efficiency and avoid queue spillover by the cooperation between traffic supply and demand. A general CCSD control framework is constructed by the control relationship description and discrete-time statespace equations. Furthermore, the uniform matrix description of CCSD is put forward under the framework to fast solve the problem by matrix calculation. An artificial intelligence planning model on CCSD is established by an objective function compromising between throughput and fairness to satisfy the control requirements under dynamic unknown traffic environment. CCSD is compared with the Webster method and capacity-aware back-pressure (CABP) control in the experiments by both the simulation data and investigation data under unsaturated, saturated, and oversaturated conditions. The results show that CCSD is superior to CABP control and the Webster method in the throughput, the number of stops, and stop time, and can avoid the queue spillover. Accordingly, CCSD can be used to improve the efficiency and avoid queue spillover at intersection under all traffic conditions.
10 January 2017 in Journal of the Optical Society of America A
Background subtraction is the fundamental step of moving object detection. As current methods use specific data-based or data learning methods to represent background, this will induce the problem of model distortion while the scenarios cannot meet their assumptions or model conditions. A novel data-driven framework is proposed to represent the background using the intrinsic characteristics of the background. In the framework, the model-free adaptive control method is used as an instance to analyze the background by its status of the nearest time instants, and linearize them dynamically from a pure data perspective. To overcome the occlusion of foreground objects, the selective update method is employed to satisfy the background update. Experiments are carried out under different video conditions to compare algorithm performance with state-of-the-art background models. The results show that the proposed method has reached over 95% in F-measure and percentage of correct classification in most cases, which is better than other state-of-the-art methods. Furthermore, the proposed method shows better robustness in severe video conditions, including bad weather and night cases, and its simplified data-driven control laws make it suitable for outdoor video surveillance.
20 July 2020 in IEEE Access
Lane-Changing Decision (LCD) behavior is complex, dangerous, and varied by the Driver’s Psychology (DP) and Driving Style (DS). For lacking the consideration of DP and DS, the existing LCD models cannot predict the LCD process varying with the drivers and traffic conditions accurately. To deal with the problems, a new LCD model coupling DP and DS, named DP&DS-LCD, is put forward. In the model, a psychological field model is constructed to represent the DP effect on the scene vehicles. And then, K-means and K-Nearest Neighbor (KNN) algorithm are respectively adopted in learning and recognition phases to recognize the current driving style pattern. Finally, based on the DP and DS, the multi-Grained Cascade Forest (gcForest) algorithm is applied to predict the LCD behavior. In experiments, DP&DS-LCD is compared with other three LCD models by using the opening I-80 database from Next Generation Simulation project (NGSIM). And the results showed that the DP&DS-LCD model achieved the best performance. Therefore, the DP&DS-LCD model is effective and could provide support for the decision of autonomous vehicles by predicting the surrounding vehicles’ Lane-Changing (LC) behavior.
8 July 2024 in IEEE Transactions on Geoscience and Remote Sensing
Using a deep learning method to detect buried pipes on ground penetrating radar (GPR) data is popular for its excellent performance potential. However, this needs a large amount of high-quality training data, which leads to time-consuming and labor-intensive data annotation work. Semi-supervised learning method provides a solution for the situation of small sample size. However, the original semi-supervised learning methods lack control over the quality of pseudo labels during model training. To address the problem, a new prior knowledge-guided semi-supervised deep learning method is proposed to improve the model performance under the small sample. In the method, a prior structure feature (PSF) is constructed to control the quality of pseudo label during semi-supervised learning. The structure response index (SRI) is designed to segment the structure subject out for eliminating disturbance from unstructured information. Then, the PSF is represented using histograms of oriented gradients (HOGs) and Fourier descriptors (FDs) to describe the structure’s edge orientation and shape. Based on the PSF, LightGBM is employed as a discriminator to screen out high-quality pseudo labels for model training. In the experiments, comparison experiments with some state-of-the-art supervised learning methods and semi-supervised learning methods were carried out to validate the method’s performance. The ablation studies on labeled dataset size, pseudo-label confidence threshold, and PSF were conducted to validate the effectiveness of the prior structural feature in controlling the pseudo label’s quality. The results show that the PSF is effective, and the proposed method outperforms the other methods.