Multi-task perception in traffic scenes requires unified representations for object recognition, region understanding, and structural reasoning. However, heterogeneous tasks such as detection and segmentation impose conflicting optimization objectives on shared features, leading to representation imbalance and reduced robustness. We propose DDL-Net, a task-balanced panoptic perception network that addresses this issue through coordinated representation learning and task-aware decoding, introducing a Channel-Rearranging Attention Interaction (CRAI) module built upon Channel Rearranging Boosting Attention (CRBA) units to reduce attention redundancy and enhance informative channel responses, improving representation of small and degraded targets. In addition, a task-specific decoding mechanism combining the Task-Guided Feature Balancing Module (TGBM) and Frequency-Aware Task-Interaction Guided Decoder (FTIG-Decoder) aligns shared features with task semantics, enabling balanced performance across detection and segmentation tasks. Experiments on BDD100K validate the effectiveness and robustness of the proposed design.

Driver attention recognition in driving scenarios is a popular direction in traffic scene perception technology. It aims to understand human driver attention to focus on specific targets/objects in the driving scene. However, traffic scenes contain not only a large amount of visual information but also semantic information related to driving tasks. Existing methods lack attention to the actual semantic information present in driving scenes. Additionally, the traffic scene is a complex and dynamic process that requires constant attention to objects related to the current driving task. Existing models, influenced by their foundational frameworks, tend to have large parameter counts and complex structures. Therefore, this paper proposes a real-time saliency Mamba network based on the latest Mamba framework. As shown in Figure 1, our model uses very few parameters (0.08M, only 0.09~11.16% of other models), while maintaining SOTA performance or achieving over 98% of the SOTA model's performance.

Intelligent vehicles need to provide rational driving strategies for assisted driving systems based on driving scenarios. Since pedestrians and vehicles are the main players in these scenarios, accurate detection and localization of pedestrians and vehicles are crucial for intelligent driving systems to make reliable decisions in dynamic environments. However, existing pedestrian and vehicle detection models often lack robustness under dynamic and complex traffic conditions, resulting in missed detections and false alarms, which pose significant safety risks. To address this problem, we categorize complex traffic scenarios into three typical challenges: long-distance, truncation, and occlusion, and focus on designing a novel enhancement stage to make the model more robust to these challenges. In this enhancement stage, inspired by human visual perception, we design a Visual Attention Module (VAM). This module can gather high-quality horizontal and vertical spatial features and efficiently interact between horizontal and vertical spatial features, enhancing the model’s perceptual ability by mimicking optic chiasm. Additionally, we use a Feature Reconstruction Module (FRM) to reduce redundant information in the feature maps and enhance the model’s inference ability. We conduct comprehensive experiments on the KITTI benchmark and Cityscapes dataset, and the experimental results demonstrate that our algorithm achieves state-of-the-art performance in various challenging scenarios.

Neural Architecture Search (NAS) is a core component of automated machine learning, enabling the automatic discovery of task-specific architectures. However, directly employing raw data for NAS faces challenges in terms of large data volume, low search efficiency, and potential data privacy concerns. To address these limitations, we condense large-scale target datasets into high information-density proxy datasets. And we propose a novel pipeline, HID-NAS, which skillfully exploits High Information Density(HID) data for efficient neural architecture search. At the algorithmic level, our algorithm focuses on the inherent problem of DARTS-derived algorithms, i.e., the discrepancy between trained supernet and selected architectures. We pioneer the phenomenon of "Selection Error" that occurs during architecture selection and introduce an attention mechanism that focuses on candidate edges to improve the accuracy of pipeline-selected architectures. Furthermore, we also formulate a regularization mechanism that employs the high information density data of the proxy dataset to transform the architectural parameter updates into an adversarial game and adjust the strengths through a normalization mechanism. These three key components – attention, regularization, and normalization – allow our pipeline to efficiently identify high-quality models using the distilled dataset. Experimental results demonstrate a significant reduction in search time, achieving high-quality models within just two minutes.

Differential Architecture Search (DARTS) has become a prominent technique for neural architecture search in recent years. Despite its merits, the issue of discretization discrepancy within DARTS still necessitates further exploration, as it can degrade in performance. In this paper, we introduce a novel algorithm termed Continuous–Discrete Alignment Optimization (DARTS-CDAO), designed to address the discretization discrepancy and thereby enhance the robustness and generalization capabilities of the discovered neural architectures. Our proposed DARTS-CDAO algorithm seamlessly integrates the discretization process into the training phase of the architecture parameters, thereby bolstering the search algorithm’s adaptability to the inherent discretization processes. Specifically, our methodology commences by formalizing the process of architecture parameter discretization. Subsequently, we introduce a coarse gradient weighting algorithm that is employed to update the architecture parameters, effectively minimizing the divergence between the representation of continuous and discrete parameters. Rigorous theoretical analysis, coupled with extensive experimental outcomes, substantiates that our proposed approach can elevate the performance of the searched models. Notably, this enhancement is achieved without incurring additional search time, rendering DARTS more robust and endowed with a heightened capacity for generalization.

With the rapid development of autonomous driving technology, the recognition/understanding of driving events has become increasingly important for improving road safety. Existing methods for recognizing driving events rely solely on the inherent features of driving scenes, lacking real-time modeling of driver attention and the integration of driver attention for understanding driving events. Research has shown that understanding driver attention will be beneficial for subsequent analysis of driving events. We propose the attention-based driving event dataset (ADED), which includes rich driving scenes, eye movement data, reasons for attention shifts, and event time windows. It enables the use of prior information about driver attention to guide the recognition of driving events. Based on our dataset, we propose a visual dual-perception network, named VP2Net, to explore the reasons behind driver attention shifts. The goal of VP2Net is to use driver attention to guide the recognition of driving events. Inspired by the human visual dual cognition process mechanism, we build a bottom-up sequential information encoding branch for extracting spatio-temporal low-level information in the driving scene. Additionally, we establish a top-down attention perceptual encoding branch that simulates the driver’s high-level visual cognitive process. It not only captures the driver’s spatial attention allocation (“where to focus”) but also performs a temporal dimensional perceptual enhancement (“when to focus”), allowing us to extract the driver’s spatial attention enhancement information. We use the driver’s spatial attention enhancement information to guide the fusion of spatio-temporal information of the driving scene and selectively highlight the core objects/areas in the current driving task/event. Finally, we compare our proposed model with other SOTA networks and visualize the results of the key components of the model. Our code is available athttps://github.com/zhao-chunyu/VP2Net

Lane detection is one of the indispensable and key elements of self-driving environmental perception. Many lane detection models have been proposed, solving lane detection under challenging conditions, including intersection merging and splitting, curves, boundaries, occlusions and combinations of scene types. Nevertheless, lane detection will remain an open problem for some time to come. The ability to cope well with those challenging scenes impacts greatly the applications of lane detection on advanced driver assistance systems (ADASs). In this paper, a spatio-temporal network with double Convolutional Gated Recurrent Units (ConvGRUs) is proposed to address lane detection in challenging scenes. Both of ConvGRUs have the same structures, but different locations and functions in our network. One is used to extract the information of the most likely low-level features of lane markings. The extracted features are input into the next layer of the end-to-end network after concatenating them with the outputs of some blocks. The other one takes some continuous frames as its input to process the spatio-temporal driving information. Extensive experiments on the large-scale TuSimple lane marking challenge dataset and Unsupervised LLAMAS dataset demonstrate that the proposed model can effectively detect lanes in the challenging driving scenes. Our model can outperform the state-of-the-art lane detection models.