Search

Interactive traffic simulation is crucial to autonomous driving systems by enabling testing for planners in a more scalable and safe way compared to real-world road testing. Existing approaches learn an agent model from large-scale driving data to simulate realistic traffic scenarios, yet it remains an open question to produce consistent and diverse multi-agent interactive behaviors in crowded scenes. In this work, we present InterSim, an interactive traffic simulator for testing autonomous driving planners. Given a test plan trajectory from the ego agent, InterSim reasons about the interaction relations between the agents in the scene and generates realistic trajectories for each environment agent that are consistent with the relations. We train and validate our model on a large-scale interactive driving dataset. Experiment results show that InterSim achieves better simulation realism and reactivity in two simulation tasks compared to a state-of-the-art learning-based traffic simulator., Comment: Accepted at IROS 2022. Author version with 8 pages, 4 figures, and 2 tables. Code and demo available at paper website: https://tsinghua-mars-lab.github.io/InterSim/

A novel method robust to outliers is proposed which integrates outliers rejection within the homography estimation pipeline with a negligible computational overhead. With a basis of algebraic criterion, the proposed method has a low level of computational time with an arbitrary large number of key-point correspondences. Given a set of 2D-to-2D matches, we formulate pose estimation problem as a low-rank homogeneous system where the solution lies on its 1D null space. Outlier correspondences are those rows of the linear system which perturb the null space and are progressively detected by projecting them on an iteratively estimated solution of the null space. Several simulations based on synthetic and real image demonstrate that the proposed method can achieve robust homography estimation to different ratios of outliers.

To Improve the robust performance of visual tracking in various kind of scene, a novel method by harmony search and co-inference learning based on multi-cues was proposed. The candidate state was achieved by the harmony search and co-inference learning. Then the harmony memory vector corresponded the biggest fitness function was chosen as the state vector. Compared with the harmony search visual tracking with the color cue and the multi-cue integrating with multiply fusion strategy, our algorithm has a better robustness in complex environments such illumination change seriously and motion blur.

Although particle filters have been widely applied from various fields, the impoverishment and degeneration of particle are important factors affecting the performance of particle filters. In this paper, harmony search was introduced into the particle resampling to enhance the algorithm. By updating the vector of harmony memory through the weight based multinomial resampling, the impoverishment and degeneration of particle was efficiently reduced. Experimental results showed that the harmony search resampling particle filters had more accurate filtering results than other typical filters such as system resampling filter.

A quadruped walking robot with the large payload capability is developed for the rescue purpose. To enhance the payload capability, the leg is designed by using the parallel mechanism. To simplify the actuation system, the actuators are selected as electric motors. Foot force is very important for the gait planning and stability control. Traditional foot force sensors are often easy to corrupt due to the strong impact from the ground. In this study, the indirect foot force sensing method is proposed, where the foot force is estimated from the motor current based on the dynamics model. To improve the accuracy of the dynamics model, an improved dynamics model is established to compensate unknown parameters and frictions by using the system identification technique. Because additional foot force or joint torque sensors are not required, the size and weight of the robot do not increase. The effectiveness of the proposed method is verified by the experiments.

To satisfy the rescue tasks, a heavy-duty quadruped robot is developed recently. To guarantee the payload capability, the leg design is based on a parallel mechanism and the actuators are selected as powerful motors. Foot force sensing is very important for the gait control. Traditional foot force sensors are easy to corrupt due to the strong repeated impact from the ground. In this study, the foot force estimation method is proposed from the distributed current sensing in the motor driver. It is based on a dynamic model of the robot. To compensate unknown model parameters and frictions, an improved dynamic model based on neural networks is estimated using the system identification technique from the measured data. No extra foot force or motor torque sensors are used, and the size and weight of the robot do not increase. The experiments are conducted to verify the effectiveness.

Advanced prediction of the geological condition of anterior rock near tunnel's work area in the hydraulic tunnel construction not only to protect the safety for the construction unit, but also can save a lot of manpower, financial and material resources. Ground Penetrating Radar, GPR for short, has been widely used for its easy operation, rapid, non-destructive, veracity, intuition and etc. In this paper, the introduction of basic working principle and the application and skills of GPR in advanced prediction of water delivery tunnel was produced.

This paper presents a novel embedded motion controller in a high precision simple automatic thyristor module die bonder against tradition PCI card based motion controller. Based on the analysis of die-bonding system structure and operation principle, a solution of 4-axes motion controller integrated with ARM Cortex-M3 microprocessor and special DSP motion control microchip is put forward. After the hardware architecture design and software programming are completed, the embedded motion controller meets the control accuracy, real-time and reliability in simple automatic die-bonding system. The innovative design not only improves product quality and production efficiency, but also reduces production costs. The controller is practical and can be widely applied in the field of industrial control.

Based on the previous undisturbed frozen field test, the research and analysis on the Size effect of fracture toughness K IC of vertical crack and herringbone crack of undisturbed frozen soils specimens of test I is performed emphatically, and it is found the rules that it varies with volume change. Meanwhile, a theoretical prediction by using the Gray System Theory methods is made and the theoretical prediction results K IC has been gotten, which is compared with the results of the test and the error fits the accuracy requirement. Thus it has been made a meaningful exploration for the research method of Size effect of fracture toughness K IC of test I of undisturbed frozen soils.

This paper designs an embedded vision processing platform for simple automatic die-bonding system by using ARM microprocessor as the vision processor, Linux as the operation system and USB camera for image acquisition equipment to complete wafer image acquisition, processing, wafer alignment and defects inspection. After the hardware platform is constructed and the software design is implemented, experimental results show that the platform has good vision processing features in real-time, simplicity and efficiency. The platform is competent for wafer image processing and analysis in die bonders and can be widely applied in the field of industry measurement and control.

A general automatic domain concept hierarchy construction procedure is presented in this paper. This is a domain independent construct a domain concept hierarchy from a domain corpus . The construction procedure mainly includes domain terminology extraction, word sense disambiguation, similarity computation, hierarchy construction and subsumption relation detection. All extracted candidate terms are ranked first, then one can select the top terms as domain terminologies. Frequency ratio and entropy of a word are considered to rank candidate terms. Relations between terms are taken into account for words in WordNet, while distributional similarity is used to compute similarity between words outside WordNet. Experiments on two domain corpus show that the proposed procedure is feasible and can get reasonable concept hierarchy.

This paper, residual high-frequency boost method is imported into the iteration process of the simulation-correction iteration image reconstruction algorithm, and produces more high-frequency details information for the algorithm, which improves the algorithm’s performance. We use Matlab to implement the algorithm and the results indicate that importing the residual high-frequency boost into the image reconstruct algorithm can highly improve the performance.

Adaptive filters are important application in the signal processing field. This paper discusses the adaptive local noise removal algorithm proposed by Professor Rafael C. Gonzalez and points out its shortcomings. To improve the algorithm, a new method, geometric mean based adaptive local noise removal algorithm, has been proposed. The simulation results indicate that the new algorithm is more satisfactory. The mean square error lscrmse is reduced by 1/4. The signals to noise ratios (i.e., SNR, SNRm, PSRN) are raised by 1/10.This algorithm has been shown promise for applications.