arXiv subjects

Xiaolin Song

Publications and source records attributed to Xiaolin Song.

12 recordsLinked to original sources

arXiv2024

Knowledge Transfer from Simple to Complex: A Safe and Efficient Reinforcement Learning Framework for Autonomous Driving Decision-Making

A safe and efficient decision-making system is crucial for autonomous vehicles. However, the complexity of driving environments limits the effectiveness of many rule-based and machine learning approaches. Reinforcement Learning (RL), with its robust self-learning capabilities and environmental adaptability, offers a promising solution to these challenges. Nevertheless, safety and efficiency concerns during training hinder its widespread application. To address these concerns, we propose a novel RL framework, Simple to Complex Collaborative Decision (S2CD). First, we rapidly train the teacher model in a lightweight simulation environment. In the more complex and realistic environment, teacher intervenes when the student agent exhibits suboptimal behavior by assessing actions' value to avert dangers. We also introduce an RL algorithm called Adaptive Clipping Proximal Policy Optimization Plus, which combines samples from both teacher and student policies and employs dynamic clipping strategies based on sample importance. This approach improves sample efficiency while effectively alleviating data imbalance. Additionally, we employ the Kullback-Leibler divergence as a policy constraint, transforming it into an unconstrained problem with the Lagrangian method to accelerate the student's learning. Finally, a gradual weaning strategy ensures that the student learns to explore independently over time, overcoming the teacher's limitations and maximizing performance. Simulation experiments in highway lane-change scenarios show that the S2CD framework enhances learning efficiency, reduces training costs, and significantly improves safety compared to state-of-the-art algorithms. This framework also ensures effective knowledge transfer between teacher and student models, even with suboptimal teachers, the student achieves superior performance, demonstrating the robustness and effectiveness of S2CD.

cs.RO↗

arXiv2023

A Delay Compensation Framework Based on Eye-Movement for Teleoperated Ground Vehicles

An eye-movement-based predicted trajectory guidance control (ePTGC) is proposed to mitigate the maneuverability degradation of a teleoperated ground vehicle caused by communication delays. Human sensitivity to delays is the main reason for the performance degradation of a ground vehicle teleoperation system. The proposed framework extracts human intention from eye-movement. Then, it combines it with contextual constraints to generate an intention-compliant guidance trajectory, which is then employed to control the vehicle directly. The advantage of this approach is that the teleoperator is removed from the direct control loop by using the generated trajectories to guide vehicle, thus reducing the adverse sensitivity to delay. The delay can be compensated as long as the prediction horizon exceeds the delay. A human-in-loop simulation platform is designed to evaluate the teleoperation performance of the proposed method at different delay levels. The results are analyzed by repeated measures ANOVA, which shows that the proposed method significantly improves maneuverability and cognitive burden at large delay levels (>200 ms). The overall performance is also much better than the PTGC which does not employ the eye-movement feature.

cs.RO↗

arXiv2022

Visual Area of Interests based Multimodal Trajectory Prediction for Probabilistic Risk Assessment

Accurate and reliable prediction of driving intentions and future trajectories contributes to cooperation between human drivers and ADAS in complex traffic environments. This paper proposes a visual AOI (Area of Interest) based multimodal trajectory prediction model for probabilistic risk assessment at intersections. In this study, we find that the visual AOI implies the driving intention and is about 0.6-2.1 s ahead of the operation. Therefore, we designed a trajectory prediction model that integrates the driving intention (DI) and the multimodal trajectory (MT) predictions. The DI model was pre-trained independently to extract the driving intention using features including the visual AOI, historical vehicle states, and environmental context. The intention prediction experiments verify that the visual AOI-based DI model predicts steering intention 0.925 s ahead of the actual steering operation. The trained DI model is then integrated into the trajectory prediction model to filter multimodal trajectories. The trajectory prediction experiments show that the proposed model outperforms the state-of-the-art models. Risk assessment for traffics at intersections verifies that the proposed method achieves high accuracy and a low false alarm rate, and identifies the potential risk about 3 s before a conflict occurs.

cs.HC↗

arXiv2022

Predicted Trajectory Guidance Control Framework of Teleoperated Ground Vehicles Compensating for Delays

Maneuverability and drivability of the teleoperated ground vehicle could be seriously degraded by large communication delays if the delays are not properly compensated. This paper proposes a predicted trajectory guidance control (PTGC) framework to compensate for such delays, thereby improving the performance of the teleoperation system. The novelty of this PTGC framework is that teleoperators intended trajectory is predicted at the vehicle side with their delayed historical control commands and the LiDAR 3D point cloud of the environment, and then the vehicle is guided by the predicted trajectory. By removing the teleoperator from the direct control loop, the presented method is less sensitive to delays, and delays are compensated as long as the prediction horizon exceeds the delays. Human-in-the-loop simulation experiments are designed to evaluate the teleoperation performance with the proposed method under five delay levels. Based on the repeated measurement analysis of variance, it is concluded that the PTGC method can significantly improve the performance of the teleoperated ground vehicles under large delays(>200ms), such as the task completion time (TCT), deviation to centerline (D2C) and steering effort (SE). In addition, the results also show that teleoperators can adapt to smaller delays, and the presented method is ineffective in such cases.

eess.SY↗

arXiv2022

S4OD: Semi-Supervised learning for Single-Stage Object Detection

Single-stage detectors suffer from extreme foreground-background class imbalance, while two-stage detectors do not. Therefore, in semi-supervised object detection, two-stage detectors can deliver remarkable performance by only selecting high-quality pseudo labels based on classification scores. However, directly applying this strategy to single-stage detectors would aggravate the class imbalance with fewer positive samples. Thus, single-stage detectors have to consider both quality and quantity of pseudo labels simultaneously. In this paper, we design a dynamic self-adaptive threshold (DSAT) strategy in classification branch, which can automatically select pseudo labels to achieve an optimal trade-off between quality and quantity. Besides, to assess the regression quality of pseudo labels in single-stage detectors, we propose a module to compute the regression uncertainty of boxes based on Non-Maximum Suppression. By leveraging only 10% labeled data from COCO, our method achieves 35.0% AP on anchor-free detector (FCOS) and 32.9% on anchor-based detector (RetinaNet).

cs.CV↗

arXiv2021

FPCR-Net: Feature Pyramidal Correlation and Residual Reconstruction for Optical Flow Estimation

Optical flow estimation is an important yet challenging problem in the field of video analytics. The features of different semantics levels/layers of a convolutional neural network can provide information of different granularity. To exploit such flexible and comprehensive information, we propose a semi-supervised Feature Pyramidal Correlation and Residual Reconstruction Network (FPCR-Net) for optical flow estimation from frame pairs. It consists of two main modules: pyramid correlation mapping and residual reconstruction. The pyramid correlation mapping module takes advantage of the multi-scale correlations of global/local patches by aggregating features of different scales to form a multi-level cost volume. The residual reconstruction module aims to reconstruct the sub-band high-frequency residuals of finer optical flow in each stage. Based on the pyramid correlation mapping, we further propose a correlation-warping-normalization (CWN) module to efficiently exploit the correlation dependency. Experiment results show that the proposed scheme achieves the state-of-the-art performance, with improvement by 0.80, 1.15 and 0.10 in terms of average end-point error (AEE) against competing baseline methods - FlowNet2, LiteFlowNet and PWC-Net on the Final pass of Sintel dataset, respectively.

cs.CV↗

arXiv2021

Haar-Weave-Metropolis kernel

Recently, many Markov chain Monte Carlo methods have been developed with deterministic reversible transform proposals inspired by the Hamiltonian Monte Carlo method. The deterministic transform is relatively easy to reconcile with the local information (gradient etc.) of the target distribution. However, as the ergodic theory suggests, these deterministic proposal methods seem to be incompatible with robustness and lead to poor convergence, especially in the case of target distributions with heavy tails. On the other hand, the Markov kernel using the Haar measure is relatively robust since it learns global information about the target distribution introducing global parameters. However, it requires a density preserving condition, and many deterministic proposals break this condition. In this paper, we carefully select deterministic transforms that preserve the structure and create a Markov kernel, the Weave-Metropolis kernel, using the deterministic transforms. By combining with the Haar measure, we also introduce the Haar-Weave-Metropolis kernel. In this way, the Markov kernel can employ the local information of the target distribution using the deterministic proposal, and thanks to the Haar measure, it can employ the global information of the target distribution. Finally, we show through numerical experiments that the performance of the proposed method is superior to other methods in terms of effective sample size and mean square jump distance per second.

stat.ME↗

arXiv2021

3rd Place Solution for VisDA 2021 Challenge -- Universally Domain Adaptive Image Recognition

The Visual Domain Adaptation (VisDA) 2021 Challenge calls for unsupervised domain adaptation (UDA) methods that can deal with both input distribution shift and label set variance between the source and target domains. In this report, we introduce a universal domain adaptation (UniDA) method by aggregating several popular feature extraction and domain adaptation schemes. First, we utilize VOLO, a Transformer-based architecture with state-of-the-art performance in several visual tasks, as the backbone to extract effective feature representations. Second, we modify the open-set classifier of OVANet to recognize the unknown class with competitive accuracy and robustness. As shown in the leaderboard, our proposed UniDA method ranks the 3rd place with 48.49% ACC and 70.8% AUROC in the VisDA 2021 Challenge.

cs.CV↗

arXiv2021

2nd Place Solution for Waymo Open Dataset Challenge -- Real-time 2D Object Detection

In an autonomous driving system, it is essential to recognize vehicles, pedestrians and cyclists from images. Besides the high accuracy of the prediction, the requirement of real-time running brings new challenges for convolutional network models. In this report, we introduce a real-time method to detect the 2D objects from images. We aggregate several popular one-stage object detectors and train the models of variety input strategies independently, to yield better performance for accurate multi-scale detection of each category, especially for small objects. For model acceleration, we leverage TensorRT to optimize the inference time of our detection pipeline. As shown in the leaderboard, our proposed detection framework ranks the 2nd place with 75.00% L1 mAP and 69.72% L2 mAP in the real-time 2D detection track of the Waymo Open Dataset Challenges, while our framework achieves the latency of 45.8ms/frame on an Nvidia Tesla V100 GPU.

cs.CV↗

arXiv2021

Non-reversible guided Metropolis kernel

We construct a class of non-reversible Metropolis kernels as a multivariate extension of the guided-walk kernel proposed by Gustafson 1998. The main idea of our method is to introduce a projection that maps a state space to a totally ordered group. By using Haar measure, we construct a novel Markov kernel termed Haar-mixture kernel, which is of interest in its own right. This is achieved by inducing a topological structure to the totally ordered group. Our proposed method, the Delta-guided Metropolis--Haar kernel, is constructed by using the Haar-mixture kernel as a proposal kernel. The proposed non-reversible kernel is at least 10 times better than the random-walk Metropolis kernel and Hamiltonian Monte Carlo kernel for the logistic regression and a discretely observed stochastic process in terms of effective sample size per second.

stat.CO↗

arXiv2020

STC-Flow: Spatio-temporal Context-aware Optical Flow Estimation

In this paper, we propose a spatio-temporal contextual network, STC-Flow, for optical flow estimation. Unlike previous optical flow estimation approaches with local pyramid feature extraction and multi-level correlation, we propose a contextual relation exploration architecture by capturing rich long-range dependencies in spatial and temporal dimensions. Specifically, STC-Flow contains three key context modules - pyramidal spatial context module, temporal context correlation module and recurrent residual contextual upsampling module, to build the relationship in each stage of feature extraction, correlation, and flow reconstruction, respectively. Experimental results indicate that the proposed scheme achieves the state-of-the-art performance of two-frame based methods on the Sintel dataset and the KITTI 2012/2015 datasets.

cs.CV↗

arXiv2018

Temporal-Spatial Mapping for Action Recognition

Deep learning models have enjoyed great success for image related computer vision tasks like image classification and object detection. For video related tasks like human action recognition, however, the advancements are not as significant yet. The main challenge is the lack of effective and efficient models in modeling the rich temporal spatial information in a video. We introduce a simple yet effective operation, termed Temporal-Spatial Mapping (TSM), for capturing the temporal evolution of the frames by jointly analyzing all the frames of a video. We propose a video level 2D feature representation by transforming the convolutional features of all frames to a 2D feature map, referred to as VideoMap. With each row being the vectorized feature representation of a frame, the temporal-spatial features are compactly represented, while the temporal dynamic evolution is also well embedded. Based on the VideoMap representation, we further propose a temporal attention model within a shallow convolutional neural network to efficiently exploit the temporal-spatial dynamics. The experiment results show that the proposed scheme achieves the state-of-the-art performance, with 4.2% accuracy gain over Temporal Segment Network (TSN), a competing baseline method, on the challenging human action benchmark dataset HMDB51.

cs.CV↗