Search

Considering the development of object detection based on deep learning framework in recent years, it has brought a new scope for multi-source fusion in the field of autonomous driving. In this paper, we propose a new architecture with a feature pyramid attention module to fuse the projected radar data and camera data. In the proposed algorithm, the detection model of YOLOv3 is employed by us and the feature pyramid module is extended with the input interface of the radar projection image and attention module. Additionally, in order to reduce the interference information from different scales of radar projected block, a new generation mechanism of radar projection images is introduced. Finally, the radar projection image is fused in feature pyramid layers with an attention module. The result shows that the proposed fusion algorithm outperforms better than image-only network for the nuScenes dataset. The code for this research will be made available to the public at: https://github.com/yuanliangxie/nuscenes_data_process.

This paper presents an effective constrained multiple model particle filtering (CMMPF) for bearings-only maneuvering target tracking. In the proposed algorithm, the process of target tracking is factorized into two sub-problems: 1) motion model estimation and model-conditioned state filtering according to the Rao–Blackwellised theorem and 2) the target dynamic system is modeled by multiple switching dynamic models in a jump Markov system framework. To estimate the model set, a modified sequential importance resampling method is used to draw the model particles, which can be restricted into the feasible area coincide with the constrained bound. To the model-conditioned state nonlinear filtering, a truncated prior probability density function is constructed by utilizing the latest observations and auxiliary variables (target spatio–temporal features), which can guarantee the diversity and accuracy of the sampled particles. The tracking performance is compared and analyzed with other conventional filters in two scenarios: 1) uniform and time-invariant sampling scenario and 2) non-uniform and sparse sampling scenario. A conservative Cramer–Rao lower bound is also introduced and compared with the root mean square error performance of the suboptimal filters. Simulation results confirm the superiority of CMMPF algorithm over the other existing ones in comparison with respect to accuracy, efficiency, and robustness for the bearings-only target tracking system, especially for the aperiodic and sparse sampling environment.

Aiming at the modeling uncertainty and nonGaussian characteristics of dynamic systems in maneuvering target tracking, a Takagi-Sugeno (T-S) modeling particle filter method via intuitionistic fuzzy decision (IFD-TSPF) is proposed. As for the Structure recognition of T-S model, the membership degree of multiple features are calculated by the ridge reduction distribution function, and then the intuitionistic fuzzy point operator is used to calculate the decision scores corresponding to the features. According to the principle of evidence decision, the appropriate premise variables are selected. Finally, the importance density function is constructed by the estimation results of the proposed T-S model, which can effectively reduce the particle degradation. Through explanation of simulation results, which show that the performance of our algorithm is better than interacting multi-model particle filter(IMMPF) and interacting multi-model Rao-Blackwellized particle filter(IMMRBPF). When the direction of the target is suddenly changed, it can effectively track the target.

In this paper, a novel multi-source vehicle detection algorithm based on deep learning is proposed. In the proposed algorithm, in order to detect the vehicle objects, a fast deep learning algorithm based on convolutional neural network (CNN) is utilized to detect the vehicle objects, and the radar is used to obtain the position information about the vehicle objects. At the same time, a coordinate transformation method from radar coordinate system to video pixel coordinate system is presented, then the video detections and radar infromation are integrated to improve the detection performance of vehicles Finally, the experiment results based on the real datasets show that the proposed algorithm is very effective for the vehicle object detection and tracking.

Particle filtering (PF) schemes are a set of simulation-based techniques relying on proposal distributions which have a crucial impact on their performance. In this paper, we introduce a novel constrained Extended Kalman particle filter (CEPF), the constrained Extended Kalman filter is used to generate the porposal distribution. The algorithm integrates the nonlinear state constraint information and the latest measurement information simultaneously into the dynamic system transition density. The proposed algorithm selects particles with higher likelihood to propagate into the next time ste by an efficient series of constraint optimization. The method can convergent theoretically and the simulation results show estimate advantage compared with other convention filters such as Extended Kalman Filter (EKF), Unscented Kalman Filer (UKF), Generic Particle Filter(GPF), particle filter with EKF proposal (EPF) and particle filter with UKF proposal (UPF).

Although visual tracking have been greatly improved in the last decade, there are still many challenges that are not fully resolved. Of these challenges, occlusions, which can be long lasting, are often ignored. Under the framework of particle filtering, this paper uses the incremental principal component analysis subspace method to learn an orthogonal subspace, then gets the linear representation of target appearance. To avoid the tracking drift, an occlusion handling scheme was proposed. In this scheme, firstly, determinate state of the target by the variance of the reconstruction error, then predict the position of the target by velocity prediction method and particle filtering method. The experiments revealed that proposed algorithm is impactful to deal with occlusion.

In order to cope with the complex variation of target appearance during visual tracking, a robust tracking algorithm based on multi-scale kernelized least squares (KLS) is proposed. First, by showing that the dense sampling set of translated patches is circulant, using the well-established theory of circulant matrices, kernelized least squares is efficient computed with fast Fourier transform (FFT). Second, we propose a method to judge whether target and background changed significantly by the response variance of all candidate targets, and use it to improve the model updating. Meanwhile, robust scale estimation based on a scale pyramid representation is proposed to handle large scale variations. Extensive experimental results show that our approach outperforms state-of-the-art methods, while operating at real-time.

To improve the performance of multi-object tracking in the complex scenario with frequent occlusions and cluttered backgrounds, a novel online multi-object tracking algorithm based on fuzzy logic is proposed. In the proposed algorithm, firstly, the similarity measure of multiple features between the objects and the measurements are calculated, including the background-weighted color feature, histogram of oriented gradients feature, local binary pattern feature and spatial distance feature. Secondly, the fuzzy rule base is constructed by incorporating the expert knowledge, which can adaptively allocate the weight of each feature by using fuzzy logic. The association probabilities between objects and measurements are substituted by the weighted sum of multiple features' similarity measure, which can effectively improve the accuracy of data association. Experimental results using challenging public datasets demonstrate that the improved performance of the proposed algorithm, compared with other state-of-the-art tracking algorithms.

For the nonlinear and non-Gaussian filtering problem of target tracking, a novel Gaussian sum quadrature particle filter(GSQPF) based on Gauss-Hermite quadrature and Gaussian sum particle filter is proposed. In the proposed algorithm, according to the advantage of Gaussian-Hermite quadrature points in the nonlinear approximation and the diversity of quadrature points, we introduce a set of quadrature point probability densities to approximate the important density function, the filtering and prediction densities are approximated as finite Gaussian mixtures. Because of the advantage of Gaussian mixture and the particle filtering, it can effectively improve the performance. The simulations show that the presented filter can outperform both Gaussian sum particle filter(GSPF) and quadrature particle filter(QPF).

In this paper, a new data association algorithm based on intuitionistic fuzzy clustering for multi-target tracking in cluttered environment was proposed. In the proposed algorithm, the joint association probabilities in JPDAF are reconstructed by utilizing the intuitionistic fuzzy membership degree of the measurement belonging to the target. In order to compute the intuitionistic fuzzy membership degree, a new intuitionistic fuzzy clustering method is proposed. At the same time, to deal with the uncertainty of the measurements, a new weight assignment is introduced. Finally, the simulation results show that the proposed algorithm is effective, and the performance of tracking is higher than the JPDAF algorithm.

In this paper, a new algorithm for multiple maneuvering target tracking is proposed. The proposed algorithm which is based on separating the multiple maneuvering target tracking into three parts-the data association, the estimation of the single target dynamic model and the estimation of the single target tracking subproblems conditional on the data association and the target dynamic model. Where the data assciation subproblem can be solved by the fuzzy data assciation, the single target dynamic model by the Rao-Blackwellized particle filter (MMRBPF) and the single target tracking by Kaiman filter or extend Kaiman filter. Finally, the experiment results show that the proposed algorithm can effectively track multiple manuvering targets.

To solve the problem of initiating tracks for multi-target in dense clutters environment, a Mean shift track initiation algorithm based on Hough transform is proposed. In the algorithm, firstly, hough transform is applied to transform observation points from input space, referred to as feature space into curves in a special parameter space; then a Mean shift clustering algorithm is executed to cluster the items gained in the parameter space, and the problem of peak seeking is also solved adaptively. Furthermore, a fuzzy influential factor, which is based on the vote number of accumulation matrix and distance between items in the parameter space and clustering center, is defined to design kernel function of Mean shift; thus clutters are removed more effectively. Experimental results show that proposed algorithm has high detection accuracy and can initiate tracks effectively.

In this paper, a new adaptive alpha - beta filter is proposed for maneuvering target tracking based on fuzzy logic. Firstly, we analyzed that the change of course angle and the measurement residual could reflect the maneuverability of the target real time, then the two parameters are used as the input variables of the fuzzy logic, and the fuzzy logic rules of the maneuverability are designed to calculate the parameters of the alpha - beta filter. The experiment results show a higher performance in maneuvering targets tracking problems and the method could be easily applied to real project.

In this paper, a new target tracking algorithm based on Gauss-Hermite quadrature is proposed in passive sensor array. Firstly, the quadrature Kalman filter (QKF) that used statistical linear regression (SLR) to linearize a nonlinear function through a set of Gauss-Hermite quadrature points is analyzed for passive target tracking. The performance of the filter is more accurate than the extended Kalman filter (EKF), the pseudo linear kalman filter (PLKF) and the unscented Kalman filter (UKF) in nonlinear dynamic system. Secondly, in order to avoid the unobservability problem of passive target tracking, a nonlinear measurement model of multiple passive sensors is founded, and the algorithm can deal with the case of non-Gaussian noise. Finally, the simulation results show that the proposed algorithm is effective, and its performance is superiority over above methods.

For maneuvering target tracking with two passive sensors in clutter environment, a novel tracking algorithm based on maximum entropy fuzzy clustering is proposed. Firstly, the interacting multiple models (IMM) approach is used to solve the maneuver problem of the target, and the false alarms generated by clutter are accommodated through maximum entropy fuzzy probabilistic data association filter (MEF-PDAF). Secondly, in order to avoid the unobservability problem of passive target tracking, a nonlinear measurement model of two passive sensors is founded. Finally, the simulation results show that the proposed algorithm has the advantages over the conventional IMM-PDAF algorithm in terms of simplicity and efficiency

Multi-target tracking is a key problem in the field of multi-sensor data fusion. A novel algorithm of multip-target tracking based on multiple FCMs data association is proposed for the multi-sensor multi-target (MSMT) tracking system. The algorithm can make full use of the information embedded in the measurements of sensors, make the minimization of the variance of the measurement fusion value through a linear weighting, and with a great reduction of the tracking errors caused by the association errors. Then the flow diagram is presented. In a scenario having five sensors, and five targets, the simulation results show that the proposed algorithm has the advantages over the existing ones of simplicity and efficiency.