FZFengyuan Zhang

Fengyuan Zhang

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Fengyuan Zhang has a total of 14 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Joint Accuracy and Deadline Evaluation of Compressed Models for Embedded Perception

    by Fengyuan Zhang, Zixuan Guo

    Compressed detectors for embedded perception are characterized by an accuracy figure and a latency figure that are measured separately, although a downstream safety function needs every individual output to be both correct and on time. This paper presents JADE, a per-sample, time-stamped joint accuracydeadline evaluation protocol that classifies each detection or warning event as correct and timely, correct but late, wrong but timely, or wrong and late, and reports the error rate, the timeout rate, their overlap, a coupling lift and the timely-correct rate (TCR). An application layer replays 2,703 pedestrian tracks from JAAD at 30 Hz through measured outcome models and latency traces, so that warning misses, false alarms and lead times are computed at the true output time. Across 135 measured runs on KITTI covering four public detectors, FP32 and INT8 precision, three input sizes, fixed and deadline-adaptive execution, clean and degraded inputs and three CPU-load conditions, the most accurate configuration (AP50 = 0.523) delivers no timely-correct output at a 100 ms deadline, whereas INT8 nano models keep a TCR of 0.48-0.51 under sustained contention. Errors and timeouts are nearly independent for fixed configurations (lift 0.99-1.10), while adaptive execution converts lateness into errors (lift 0.63-0.80, error rate +0.18 under contention). Perception TCR predicts the timely-correct warning rate far better than AP50 (Kendall τ = 0.66 versus 0.19), and evaluating on 2 Hz keyframes instead of 30 Hz annotations scrambles the configuration ranking (τ = 0.32).

  • Open Access

    Scenario-balanced Calibration for Post-training Quantization of Edge Perception Models

    by Fengyuan Zhang, Wenbin Shang

    Static post-training quantization (PTQ) of edge perception models fixes every activation range from a small calibration set, which in practice is drawn at random from ordinary scenes. We study how a fixed budget of calibration frames should be composed so that an INT8 detector serves both ordinary and degraded scenes. We show that with MinMax calibration the integer model depends on the calibration set only through per-tensor extremes, and propose Scenario-Balanced Calibration (SBC), which combines activation-range signatures computed with the converter's own calibration graph, a label-free FP32/INT8 disagreement probe, sensitivity-driven quotas with a coverage floor, and facility-location selection within each scenario. Experiments on YOLOX-Tiny and YOLOv8n with ONNX Runtime INT8 and 15 test conditions, including held-out degradation types, held-out severities and compound degradations, show that scenario-aware calibration consistently outperforms the clean-only default. SBC reduces the mean INT8 accuracy gap of YOLOv8n from 1.52 to 1.02 mAP points, raises the worst-condition mAP50 of YOLOX-Tiny from 14.7% to 16.4%, and adds less than 0.1 s of selection time. Ablations identify representative selection as the decisive ingredient (+0.83 points over random selection under identical quotas), show that unconstrained activation coverage collapses onto a single scenario, and confirm that random calibration sets lose accuracy as they grow, as predicted by the extreme-driven range mechanism.

  • Open Access

    A Reproducible Benchmark for On-Device Next-App Prediction and Predictive Preloading

    by Wenbin Shang, Fengyuan Zhang

    Predicting which application a smartphone user will open next allows the system to keep or start that application in memory ahead of time, turning a cold launch into a warm one. Existing studies evaluate next-app predictors with incompatible preprocessing and with accuracy metrics that ignore the memory manager, which leaves open how much launch latency a better predictor actually buys. We present a reproducible benchmark built on the public LSApp trace (292 users, 3.66 million events) that couples leakage-free prequential prediction with a tracedriven preloading simulator modeling an LRU resident set, perapp cold and warm latencies, preload timeliness, wasted preloads and cache pollution. Three findings emerge. First, standard accuracy is a weak proxy for preloading value: the proposed cache-complementary accuracy (CCA), which scores predictions only on launches that miss the resident set, orders twelve predictors by their preloading benefit almost perfectly (Spearman ρ = 0.99 versus 0.79 for Hit@1), and among personalized predictors it correlates with latency gains at r = 0.89, where Hit@1 reaches r = 0.05. Second, naive top-k preloading is counterproductive: top-1 preloading wastes 81.3% of its preloads and increases mean launch latency by 3.9%, whereas the proposed eviction-aware utility-gated policy (EUG) lowers it by up to 8.0% and dominates confidence thresholding along the cost-benefit frontier. Third, a 1,057-parameter incremental-state ranker (ISR) built on personal on-device state outperforms recurrent and selfattentive predictors and is within 0.1 points of a gradient-boosted ranker in Hit@1 (Hit@1 0.531, MRR 0.681) with a 5.2 KB ONNX graph and a 112 µs end-to-end decision latency. A streaming ONNX Runtime replay of the raw event log reproduces the offline benchmark exactly at the feature level and to within 8.8×10-7 in probability.

  • Open Access

    Hardware-Aware Co-Design of Vision-Language-Action Models for Real-Time Autonomous Driving on Edge NPUs HACT-VLA: Query-Conditioned Concept Routing with Low-Rank Integer-Friendly Fusion

    by Fengyuan Zhang

    Vision-language-action (VLA) models are attractive for autonomous driving because language can expose intent, support human interaction, and organize scene reasoning, yet contemporary VLA pipelines are difficult to deploy under the deterministic latency, memory, and operator constraints of automotive edge neural processing units (NPUs). This paper presents HACT-VLA, a concrete hardware-aware co-design that replaces unrestricted cross-modal attention in the action head with a query-conditioned concept-token route, a low-rank multiplicative fusion operator, and an integerfriendly decision path. The key idea is to exploit the structure of driving decisions: a language action query such as "may I turn left?" does not require every semantic concept inferred from the scene. HACT-VLA predicts 21 interpretable driving concepts, scores them with a compact language-conditioned router, selects a fixed top-𝑘 subset, and evaluates the requested action using only dense linear projections, elementwise operations, and a small top-𝑘 primitive. This yields a fixed-shape computation graph suitable for ahead-of-time deployment while reducing the trainable action-head parameter count by 25.2% relative to a dense query baseline in our implementation. The study is organized around the public BDD-OIA task, which contains four driving actions and 21 action-inducing explanation concepts. We provide a verified real-data pipeline for the public 2048-dimensional Faster-RCNN-derived BDD-OIA embeddings and a fully measured BDD-OIAcompatible surrogate experiment for hardware/co-design analysis. On the executed surrogate, the dense query model reaches 0.9998 macro-F1 with 133.7k parameters; HACT-VLA uses 100.0k parameters and reaches 0.9964 macro-F1 at 𝑘 = 4. The same model remains above 0.9938 macro-F1 under additive feature noise with standard deviation 0.6. Sensitivity experiments reveal a non-monotonic routing effect: increasing 𝑘 from 4 to 12 reduces macro-F1 from 0.9964 to 0.9896, indicating that hardware-driven sparsity can also act as semantic regularization. Actual CPU timing further shows that naive dynamic INT8 is not automatically faster for a tiny batch-1 policy head because quantization dispatch overhead dominates; this finding motivates static NPU compilation rather than assuming low precision guarantees low latency. All reported timing values are measured on the provided runtime, and no CPU number is relabeled as NPU latency. The resulting method, analysis, and code form a reproducible basis for completing device-specific NPU deployment without changing the learned policy interface.

  • Top Cited Papers
    Open Access

    Adaptive Sensor Fusion for Robust Perception in Dense Fog: A Gated Vision and LiDAR Integration Framework

    by Fengyuan Zhang, Zixuan Guo, Jianbo Ding, Jingyun Yang, Wenhe Liu

    Autonomous driving systems face critical perception failures in dense fog, where conventional RGB cameras suffer from severe degradation due to atmospheric scattering and reduced visibility. This paper presents an adaptive multi-modal fusion framework that synergistically integrates gated imaging with 3D LiDAR point clouds to achieve robust obstacle detection under visibility conditions as low as 50 m. Unlike standard cameras that passively capture scattered ambient light, gated cameras employ time-synchronized active illumination to physically filter backscattered photons, preserving structural features even in low-visibility scenarios. We propose a novel Adaptive Feature-Weighting Network (AFW-Net) that dynamically adjusts sensor modality contributions based on real-time environmental degradation assessment. The framework incorporates three key innovations: (1) a cross-modal feature extraction module that exploits the complementary physical properties of gated imaging and LiDAR, (2) an attention-based adaptive fusion mechanism that quantifies per-modality reliability through uncertainty estimation, and (3) a degradation-aware training strategy using weather-specific augmentation. Extensive experiments on the Princeton Automated Driving Dataset demonstrate that our approach maintains detection average precision (AP) above 82% under dense fog conditions (50 m visibility), representing a 23.7% improvement over state-of-the-art RGB-LiDAR fusion methods that exhibit substantial performance degradation to 58.4% AP. Ablation studies validate the necessity of each component, and cross-dataset evaluation confirms the generalization capability of the proposed framework. The adaptive weighting mechanism proves particularly effective, dynamically rebalancing modality contributions across the gated imaging and LiDAR branches while maintaining LiDAR geometric constraints. This work establishes a robust perception paradigm for safety-critical autonomous systems operating in low-visibility environmental conditions.


  • Top Cited Papers
    Open Access

    DRL-Adapt: Deep Reinforcement Learning for Adaptive Routing Convergence Optimization in Large-Scale Networks

    by Boyuan Wang, Zimeng Wang, Wenyu Zhao, Fengyuan Zhang, Wenbin Shang

    Large-scale fully-meshed networks often experience prolonged routing convergence after failures because routing protocol parameters are statically configured and cannot adapt to changing network conditions. To address this limitation, this paper proposes DRL-Adapt, a deep reinforcement learning-based framework that dynamically tunes routing protocol parameters using network event-stream observations. The framework employs a dueling prioritized deep Q-network agent to jointly adjust Hello/Dead interval timers and routing weights according to real-time network states. We evaluate the proposed method using a high-fidelity simulator constructed from CAIDA topologies under diverse failure scenarios and network scales. Experimental results demonstrate that DRL-Adapt significantly accelerates recovery and reduces control overhead compared with conventional static configurations. On average, the proposed approach decreases fault convergence time by 32.7% and lowers control-plane overhead by 24.3%, while preserving routing stability. Moreover, the learned policy generalizes across networks ranging from 50 to 500 nodes and incurs minimal inference latency ($< $ 15 ms/decision), indicating practicality for practical deployment.