Yurong You

The objective of this study was to create a framework for training and selecting machine learning algorithms (MLA) to classify cow health status daily using data from multiple automated health monitoring systems (AHMS), including wearable and nonwearable sensors, combined with nonsensor data of potential value for predicting cow health. The work presented in this manuscript is part of a series of studies aimed at identifying a single candidate algorithm that, upon extensive refinement and further development, could be deployed in a commercial dairy operation to identify cows potentially affected by health disorders for clinical examination. Data from AHMS and other cow features and performance data, including the clinical health status of cows, were collected in a prospective cohort study including Holstein cows (n = 1,252). Data from AHMS used for MLA training included rumination, eating, and physical activity measured in the neck (neck sensor), temperature and physical activity measured in the reticulorumen (bolus sensor), physical activity and resting measured in the leg (leg sensor), and milk yield, milk electrical conductivity, and milk components (parlor sensors). Other non-AHMS data used were temperature and humidity index, cow and calving event features, and current and previous lactation performance and management indicators. The dataset included 22,415 cow-day records with 49 features. The dataset was split into training and testing sets in an 80:20 ratio, resulting in 17,887 and 4,528 cow-day records, respectively. Data imputation and standardization were applied automatically or manually. A diverse set of nondeep learning (n = 26) MLA were trained and compared using the open-source automated ML (AutoML) tool Lazy Predict Classifier (LZP). Upon selection of the best-performing nondeep learning algorithms (i.e., XGBoost, AdaBoost, Nearest Centroid, and Bernoulli Naive Bayes) from the pool tested with LZP, classifier algorithms were compared with more complex deep learning algorithms (multilayer perceptron, recurrent neural networks, long short-term memory networks, and gated recurrent unit models) not included in LZP. All algorithms underwent training and evaluation before selection of a single best-performing algorithm, using several metrics of performance. Ensemble learning models, particularly XGBoost, achieved the best performance and balanced results with a sensitivity of 82.4% and a precision of 42.6% combined with a specificity of 86.4% and a negative predictive value of 97.6%. This model also had the highest F1-score (0.56) and area under the curve (84.4%). The XGBoost algorithm also demonstrated robustness in handling missing data. Our comprehensive approach to MLA screening and selection enabled informed decisions in selecting a suitable algorithm for identifying cows for clinical examination based on the daily prediction of health disorder occurrence. The combination of the AutoML tool LZP and manual refinement and testing of multiple MLA provided a robust framework for comparing multiple ML models. Ensemble classification learner algorithms such as XGBoost and Adaboost might outperform other deep learning and nondeep learning algorithms for classifying cow health daily using AHMS and other cow management and performance indicators.

Autoregressive Transformers are increasingly being deployed as end-to-end robot and autonomous vehicle (AV) policy architectures, owing to their scalability and potential to leverage internet-scale pretraining for generalization. Accordingly, tokenizing sensor data efficiently is paramount to ensuring the real-time feasibility of such architectures on embedded hardware. To this end, we present an efficient triplane-based multi-camera tokenization strategy that leverages recent advances in 3D neural reconstruction and rendering to produce sensor tokens that are agnostic to the number of input cameras and their resolution, while explicitly accounting for their geometry around an AV. Experiments on large-scale AV datasets and a state-of-the-art neural simulator demonstrate that our approach yields significant savings over current image patch-based tokenization strategies, producing up to 72% fewer tokens, resulting in up to 50% faster policy inference while achieving the same open-loop motion planning accuracy and improved offroad rates in closed-loop driving simulations.

The rapid development of 3D object detection systems for self-driving cars has significantly improved accuracy. However, these systems struggle to generalize across diverse driving environments, which can lead to safety-critical failures in detecting traffic participants. To address this, we propose a method that utilizes unlabeled repeated traversals of multiple locations to adapt object detectors to new driving environments. By incorporating statistics computed from repeated LiDAR scans, we guide the adaptation process effectively. Our approach enhances LiDAR-based detection models using spatial quantized historical features and introduces a lightweight regression head to leverage the statistics for feature regularization. Additionally, we leverage the statistics for a novel self-training process to stabilize the training. The framework is detector model-agnostic and experiments on real-world datasets demonstrate significant improvements, achieving up to a 20-point performance gain, especially in detecting pedestrians and distant objects. Code is available at https://github.com/zhangtravis/Hist-DA.

Monocular image-based 3D perception has become an active research area in recent years owing to its applications in autonomous driving. Approaches to monocular 3D perception including detection and tracking, however, often yield inferior performance when compared to LiDAR-based techniques. Through systematic analysis, we identified that per-object depth estimation accuracy is a major factor bounding the performance. Motivated by this observation, we propose a multi-level fusion method that combines different representations (RGB and pseudo-LiDAR) and temporal information across multiple frames for objects (tracklets) to enhance per-object depth estimation. Our proposed fusion method achieves the state-of-the-art performance of per-object depth estimation on the Waymo Open Dataset, the KITTI detection dataset, and the KITTI MOT dataset. We further demonstrate that by simply replacing estimated depth with fusion-enhanced depth, we can achieve significant improvements in monocular 3D perception tasks, including detection and tracking.