271 results about "Smoothing" patented technology

The invention relates to the technical field of image data processing, in particular to a shell defect detection method and system based on machine vision, and the method comprises the steps: collecting a to-be-detected shell image, and carrying out the smoothing of a to-be-detected shell image through an improved Gaussian filtering algorithm, and obtaining a corrected image; carrying out edge detection on the corrected image, and if the number of detected edge contours is greater than a set threshold value, judging that flaws exist; wherein in the improved Gaussian filtering algorithm, the Gaussian weight is positively correlated with the noise calibration degree of the pixel point, and is inversely correlated with the mean value of the pixel gradient amplitudes in the set window; the noise calibration degree is in positive correlation with the pixel noise possibility degree and the gray variance in the window and is in inverse correlation with the noise possibility degree in the window, and the noise possibility degree represents the defect degree. According to the method, the problems that details are lost and noise and defects are inaccurately distinguished due to excessive smoothness when an existing Gaussian filtering algorithm is used for shell detection are solved.

The invention relates to the technical field of multi-modal time series data processing, in particular to a multi-scale multi-modal fusion time series classification model, which comprises the following steps of: performing timestamp unification, missing value filling and normalization processing on input multi-modal time series data to generate standardized data; extracting long-period and short-period features by adopting a time sequence convolutional network and a one-dimensional convolutional network respectively, and performing feature alignment and fusion; cross-modal dynamic coupling is realized through cross attention and a gating mechanism, and feature noise reduction is performed in combination with adaptive threshold filtering; feature weighted fusion is completed based on multi-level feature division and a channel attention mechanism, and a comprehensive time sequence feature vector is generated; and finally, through nonlinear feature enhancement, Softmax probability prediction, sliding window smoothing processing and majority voting, outputting a classification result aligned with a timestamp. The method has the advantages of being high in feature fine granularity, good in cross-modal fusion effect and the like, and is suitable for the fields of intelligent manufacturing, behavior recognition, medical monitoring and the like.

The invention relates to the technical field of value accounting, and discloses an ecological system total production value accounting method based on GIS spatial analysis, and the method comprises the steps: carrying out the accounting of regional basic data and satellite image data, and constructing a bidirectional dynamic semantic mapping library corresponding to land change investigation land parcel classification-ecological system type; internet of Things monitoring data and statistical report data are obtained, and a standardized data set is obtained through space-time fusion of spatial interpolation-boundary correction and a sliding window smoothing method; constructing an urban area-typical district double-level accounting index system, and synchronously calculating a real object quantity and a value quantity; a model correction factor is called based on a GIS, indexes are quantified to grid units, a result is obtained through neighborhood analysis, and after parameters are calibrated, a multi-year ecological value trend map is generated in combination with the GIS; and outputting a customized result according to scenes such as ecological compensation, and outputting a final result after the customized result is qualified through three-level quality control verification. According to the method, the requirements of practical affairs such as ecological compensation and EOD projects on accounting precision and practicability can be met.

The invention relates to the technical field of path control, in particular to a combined path and laser processing planning method based on machine vision, which comprises the following steps: acquiring an image by an industrial camera to extract contour measurement parameters to identify defective materials, setting path start-stop optimization connection smooth obstacle avoidance based on a feature table, and detecting a dynamic adjustment range of a hot area. And according to the adaptive group setting threshold evaluation capability, the delay prediction adjustment power speed is calibrated, an instruction set is called to compensate, predict and correct error dynamic matching, and an optimal processing path and a synchronous processing setting table are output. According to the method, the workpiece image is obtained in real time, the machining features are analyzed, the machining precision and efficiency are improved, fine control in the machining process is achieved, effective cooperation of path planning and machining parameters is ensured, the problem of mismatching of path planning is avoided, the response precision and stability in the machining process are improved, and the machining precision and efficiency are improved. The processing quality and time management are remarkably improved, and low efficiency and quality fluctuation caused by manual adjustment in the prior art are avoided.

The invention provides a gust front wind shear identification method based on artificial intelligence, and the method comprises the steps: carrying out the noise filtering, missing value supplementary measurement, data smoothing, wind shear value calculation and sample screening extraction of collected radial speed data, and obtaining a gust front wind shear sample; performing coordinate system conversion, sample set division and data annotation on the basis of gust and front wind shear samples to obtain an expanded data set; designing and training a Mask R-CNN model architecture to obtain a gust and front wind shear identification model; and inputting the expanded data set into the gust and front wind shear identification model to carry out gust and front wind shear detection. According to the method, dependence on reflectivity factor data can be reduced, an identification model is constructed based on gust front radial speed data, gust front wind shear can be accurately identified, pixel-level segmentation and positioning of a wind shear area can be realized, and identification efficiency is improved.

The invention discloses an airborne laser radar sounding data processing method and system, and relates to the technical field of marine surveying and mapping and earth observation, and the method comprises the following steps: collecting original full-waveform data of an airborne laser radar sounding system, the original full-waveform data comprising infrared laser channel data and blue-green laser multichannel data; joint denoising processing is carried out on the original full-waveform data to obtain denoised waveform data, and the joint denoising processing comprises background noise modeling and removing based on machine learning and random noise filtering based on frequency domain low-pass filtering; a collaborative joint denoising strategy is formed by fusing background noise modeling based on machine learning and frequency domain low-pass filtering based on signal-to-noise ratio optimization. According to the method, complex background noise can be accurately estimated and removed, random noise can be adaptively filtered out, weak underwater echo signals are reserved to the maximum extent, and the problem of water depth information loss caused by excessive smoothness is effectively avoided.

The invention discloses an industrial defect detection method based on self-supervised pre-training and feature space generation. Firstly, features are extracted and mapped to a unified potential space; secondly, introducing a SimMIM framework, which is one of mainstream technologies in the current industrial vision pre-training field, to carry out self-supervised pre-training to improve the representation capability, and using an OfficientForme network to improve the reasoning speed and reduce the memory demand; secondly, performing abnormal synthesis in a feature space by adopting a generative adversarial network and enhanced Perlin noise, outputting a feature increment and a soft mask by a generator, and performing linkage updating with a pixel-level mask and an image-level label; thirdly, a segmentation-classification double-head framework is adopted, a segmentation head outputs a pixel-level anomaly graph, and a classification head outputs an image-level anomaly score; and finally, training optimization is carried out through strategies such as abnormal graph up-sampling smoothing, a grouping learning rate, multi-stage scheduling and the like. The method realizes accurate and controllable synthesis of the abnormal region under the scene without, with or with mixed supervision, and significantly improves the robustness and real-time performance.

The invention discloses a GRACE and Swarm time-varying gravity field fusion filtering method based on a state space model, and the method comprises the steps: taking a spherical harmonic coefficient as a state quantity, constructing a random walk process equation, introducing three types of observations, employing a quantization parameter for observation noise and process noise covariance, constructing according to an order / power law, and carrying out the self-adaptive updating along with the monthly; a Nelder-Mead method is adopted to search for spectral index parameters, and an EM algorithm and a statistical method are utilized to update other parameters in a closed / quasi-closed mode; obtaining the state and posterior covariance of a full time sequence by using Kalman filtering and RTS smoothing; in the GRACE and GRACE-FO window period, continuous reconstruction is carried out by means of a process model and Swarm; and outputting quality evaluation information including monthly gravity field coefficients, posterior covariance, innovative variance ratio, residual whitening test, space power spectrum, uncertainty band and the like. According to the method, while physical rationality and calculation feasibility are ensured, a continuous and stable monthly time-varying gravitational field sequence with quantifiable uncertainty is realized.

The embodiment of the invention provides a dynamic tactile rendering method based on a two-dimensional image and a finger motion track. The method comprises the following steps: acquiring an original two-dimensional image; converting a target object in the original two-dimensional image into a three-dimensional structure point cloud; respectively mapping the three-dimensional structure point cloud and the predicted complete trajectory to a unified coordinate system to obtain a corresponding point cloud data point set and a trajectory data point set; performing interpolation smoothing on the matching point set to construct a continuous geometric profile curve; calculating a first-order derivative of each section point in the geometric profile curve, and taking the first-order derivative as a stimulation value of the virtual tactile sense; converting the stimulation value into a driving voltage signal, and controlling the operation of the electrostatic adhesion tactile feedback device through the driving voltage signal; according to the dynamic tactile rendering generation method based on the two-dimensional image and the finger motion trail, the limitation that traditional tactile feedback depends on a static template and special sampling data is broken through, and automatic modeling and real-time response from two-dimensional visual input to physical tactile output are achieved.

The invention discloses a three-dimensional contour reconstruction method based on phase smoothing preprocessing in a phase shift profilometry, a storage medium and equipment. The method comprises the following steps: establishing a mathematical model between a phase shift fringe image and a phase in the phase shift profilometry to obtain an analytic expression for calculating the phase, a fringe background component and a fringe modulation amplitude; separating a modulation component from the phase shift fringe image, calculating a modulation degree of the phase shift fringe image, normalizing the modulation component according to the modulation degree, and performing low-pass filtering on the normalized modulation component; and finally, synthesizing the original background component and the modulation component subjected to low-pass filtering into a new phase shift fringe image, completing preprocessing, and calculating a phase by using the preprocessed image. According to the method, the influence of background component noise, the surface reflectivity of the object to be measured and other factors on the phase smoothing effect is avoided, the high-frequency characteristics in the phase can be reserved, the good phase smoothing effect is achieved, and the precision of the reconstructed three-dimensional contour is improved.

The application discloses a kind of big data-based optimization analysis method and system of territorial space planning, it is related to big data and optimization analysis technical field of territorial space planning, including real-time data acquisition and dynamic preprocessing, dynamic regional division, multi-model integrated space-time prediction, multi-objective dynamic optimization planning, real-time feedback and adaptive adjustment.The big data-based optimization analysis method of territorial space planning provided in the application adopts space-time weighted fuzzy clustering method, divides the territorial space into several sub-regions according to the spatial coordinates and time attributes of data, constructs the weighted distance regular objective function between data points and regional center, and updates the membership and the position of regional center iteratively, realizes the self-adaptation and space-time smoothing of regional segmentation, effectively captures the dynamic change characteristics in region.

The invention relates to the technical field of spectrum baseline correction, and provides a spectrograph baseline correction method and system based on improved empirical mode decomposition, and the method comprises the following steps: employing a sliding window smoothing method, and carrying out the preprocessing of obtained measurement spectrum data; an improved empirical mode decomposition algorithm is adopted, fitting is carried out on the preprocessed measurement spectrum data to obtain baseline data, and the improved empirical mode decomposition algorithm adopts least square fitting to construct upper and lower boundary envelopes of the spectrum data; subtracting the baseline data obtained by fitting from the obtained measurement spectrum data to obtain the measurement spectrum data after the baseline is removed; and sequentially carrying out abnormal point detection and removal, spectral power calibration and wavelength calibration on the measured spectral data after baseline removal to obtain spectral data after baseline correction. The spectral baseline is identified by improving empirical mode decomposition, the stability and boundary adaptability of mode decomposition are enhanced, and a more accurate baseline fitting result is obtained.

The invention discloses a multi-channel measurement time difference fusion method for Kalman filtering, and the method comprises the steps: obtaining the historical data of multi-channel measurement time differences, and obtaining a series of historical mean values; monitoring multi-path measurement time difference input, judging interruption when any path of judgment observation value is missing or invalid based on a historical mean value, performing interruption compensation, and dynamically adjusting process noise covariance according to interruption duration; based on historical data, dynamically distributing weights; and fusing the multiple paths of time difference data through Kalman filtering, and outputting a smoothed time difference measurement result. Through an interruption compensation strategy (virtual observation and weight adjustment) and sliding window smoothing, even if single-path or multi-path input is interrupted, continuous fusion metering time difference can still be output; the dynamic weight distribution mechanism makes full use of statistical correlation of multi-path time difference, and the interrupt path dynamically adjusts the weight according to real-time reliability; and dynamically updating the process noise covariance Q so as to adapt to a time-varying noise environment and realize adaptive noise processing.

The invention relates to the technical field of vehicle suspension system control, and discloses a multi-source data fused shock absorber self-adaptive regulation and control method and system.The method comprises the steps that multi-source signal data are synchronously collected through a sensor, and a preliminary signal set is generated in combination with environment change data; analyzing vibration signal peak amplitude and displacement data integral variation trend, distinguishing high-frequency small-acceleration and low-frequency large-amplitude pavement excitation characteristics, and determining signal contribution degree distribution; dynamically adjusting the signal weight and introducing temperature compensation to correct stiffness drift to generate a weighted vector; carrying out weighting processing, smoothing, noise suppression and nonlinear amplitude estimation to generate fusion data, and optimizing the weight to obtain a refined weighted vector; after secondary fusion, a cross-frequency-band transition smoothing technology and a least square method are adopted to determine damping characteristic parameters; according to the vehicle suspension system self-adaptive regulation and control method and system, the self-adaptive regulation and control accuracy and the working condition adaptability of the vehicle suspension system are improved.

The invention discloses a geographic information survey data management system and method applying a GIS technology, and relates to the field of geographic information survey management. Multi-source information data collected by geographic information survey equipment is preprocessed to obtain a time sequence data structure and a transition data structure which can be spliced; acquiring elevation data at the splicing position of the target time sequence data structure and the adjacent time sequence data structure, calculating and adjusting the relative position relation between the target time sequence data structure and the adjacent time sequence data structure by using the splicing transition model according to an elevation data difference; determining the width of a transition region based on the adjusted elevation data difference at the splicing position, and smoothing the transition region through the splicing transition model; splicing the transition region after smoothing processing with the time sequence data structure, and completing smooth transition at the splicing position of the target time sequence data structure and the adjacent time sequence data structure; through the transition data structure and the processing logic thereof, the accuracy of the spliced digital terrain three-dimensional model is ensured to the greatest extent.

The invention discloses a deep sea mineral resource segmentation method and device based on dynamic anchor points and iterative optimization, and relates to the field of computer vision, and the method comprises the steps: obtaining sonar point clouds and laser radar point clouds of deep sea mineral resources, combining the sonar point clouds and the laser radar point clouds into unified point clouds, and obtaining high-definition image texture features; fusing the initial geometric features and the texture features of the unified point cloud to obtain initial multi-modal features; constructing a kernel point, calculating a local structure feature and obtaining a multi-scale feature; using a graph attention network to extract global features and performing clustering to obtain class anchor points; fusing the preliminary multi-modal features, the texture features and the global features to obtain a final feature representation, calculating the similarity between the final feature representation and the class anchor points, and performing classification to obtain a preliminary segmentation result; boundary smoothing and texture correction are carried out on the preliminary segmentation result, and then fusion with multi-scale features is carried out; suspicious points are detected, error correction is carried out on the suspicious points, and segmentation is completed. According to the method, multi-modal feature iterative optimization, misclassification point continuous correction and class anchor point timely updating are realized.

The application discloses a structure point cloud data multi-scale filtering method considering environmental dynamic influence, which firstly adopts a 'K' nearest field method considering structure dynamic influence to perform dynamic filtering processing on original point cloud data, removes scattered outliers in the original point cloud data, and then divides the point cloud data into flat regions and mutation regions based on an improved principal component analysis algorithm (LMSR-PCA), adopts statistical filtering based on local surface fitting for the flat regions, and adopts spatial adaptive bilateral filtering for the mutation regions. The application is used to solve the problems that the existing point cloud data processing method cannot efficiently process massive point cloud data with complex curve characteristics, and cannot solve the problem of excessive smoothing caused by the loss of edge features in the noise reduction process of point cloud data with complex curve characteristics.

The proposed technique introduces embodiments of a computer-implemented method for interpreting image delineations as vector objects and topological extraction from segmentation by visual computational methods applied to sections (slices) of seismic volumes, in order to aid geological interpretation and sampling of parameters originating from the fracture network and its topology. Embodiments of a developed method integrates a software / application that allows the loading and generation of statistical data related to the fracture network while maintaining georeferencing and scale of the two-dimensional input data. In addition, a fracture segmentation method is shown that uses pyramid image smoothing (decomposition into hierarchical levels of resolution) in order to reduce the amount of details and aid the identification of main faults or fractures. The segmentation after this smoothing is based on adaptive thresholding segmentation.

The invention relates to a method and device for controlling the nanometer morphology of a double-sided thinned silicon wafer, and the method comprises the steps: obtaining the warping degree data of each position on the double-sided thinned silicon wafer, and obtaining a warping degree data curve; performing multi-stage differential smoothing processing on the warping degree data curve to obtain morphology parameters of the silicon wafer after double-sided thinning; and comparing the morphology parameters of the silicon wafer after double-sided thinning with a preset morphology parameter threshold value, and adjusting parameters of double-sided thinning equipment according to a comparison result. Batch production of defective products can be avoided.

The invention provides a wind profile radar spectrum peak estimation method based on a three-parameter dynamic cost function, and the method comprises the steps: 1, carrying out the preprocessing of distance and Doppler spectrum data, and carrying out the three-point moving average smoothing; step 2, noise level estimation based on statistics is carried out; step 3, carrying out candidate spectrum peak selection based on space Doppler window smoothing; 4, performing dynamic weight calculation based on covariance matrix eigenvalue decomposition; step 5, constructing a three-parameter cost function of a distance spectrum peak value RSP power item, a continuity item and a spectrum width consistency item, and performing optimization; and step 6, outputting the optimal spectrum peak trajectory to form a complete wind speed profile trajectory. According to the method, the spectrum peak identification precision and stability in a complex environment can be improved, and the u and v wind correlation coefficients of the estimation method are verified to reach 0.880 and 0.879 through a data set disclosed by an atmospheric radiation measurement website.

The invention discloses a plane alignment element segmentation and node identification method and system based on curvature characteristics, and the method comprises the steps: carrying out the smoothing of the curvature of point cloud data of a road center line through a local weighted regression algorithm, and obtaining a smooth curvature; traversing the smoothed curvature by adopting a sliding window, calculating a variable coefficient of the curvature in each window, judging the consistency of the curvature and the linear element type according to the size of the variable coefficient to obtain a preliminary candidate node, respectively extracting the curvatures of the left side and the right side of the candidate node by taking the node as a center, and carrying out least square method linear fitting to obtain a final candidate node; and obtaining an intersection point of the fitting straight lines, and determining a final node position. The method can effectively solve the problems of large curvature fluctuation of the road center line and high node identification difficulty, has the advantages of high node positioning accuracy, strong adaptability and high calculation efficiency, and can be widely applied to the fields of road digital modeling, traffic engineering design and maintenance and the like.

The invention relates to the field of data analysis, in particular to a multi-source monitoring data-oriented dynamic trend assessment method, which comprises the following steps of: acquiring and preprocessing multi-source monitoring data to obtain a historical data sliding window and a basic smoothing coefficient for trend assessment; performing weighted correlation analysis on the disturbance variable and the measured variable change sequence to obtain a working condition response decoupling factor; obtaining a trend stability factor by evaluating the geometric morphology of the smooth trajectory of the measured variable; performing working condition response and trend stability combined correction on the basic smoothing coefficient to obtain a dynamic smoothing coefficient; a smooth trend value and a heat exchanger sub-health early warning signal are obtained by performing exponential weighted moving average and change rate judgment on a measured variable, and the problem that a fixed parameter EWMA cannot distinguish a working condition adjustment response and a tiny fault trend under a multivariable complex working condition is solved.

The invention relates to an intelligent optimization method for sludge deep dehydration process parameters. Key parameters are collected in real time through an Internet of Things architecture, a multi-dimensional dynamic working condition data set is constructed by using a sliding time window and standardization processing, data quality is improved through abnormal value filtering and noise smoothing, a CNN-GRU hybrid model is input, and precise prediction of a working condition evolution trend is realized. On the basis of a prediction result, a trend guide feedback channel, a preference evolution feedback channel and a stability feedback channel are constructed, a multi-objective optimization weight is adjusted in a self-adaptive mode through normalization and weighted fusion, and a high-correlation optimal process parameter combination and closed-loop feedback control are obtained by combining NSGA-II algorithm multi-objective optimization and trend alignment degree screening. According to the scheme, the data perception and multi-target adaptive optimization capability of the sludge deep dehydration system is effectively enhanced, and the operation efficiency, the energy consumption utilization and the system stability are improved.

The application discloses a kind of intelligent anchor rod system early warning method and threshold determination method, belong to anchor rod intelligent early warning field, including S1: monitoring data preprocessing, S2: data trend regression analysis and S3: early warning method and index;Step S1 includes substep, S11: detection and correction of singular point, S12: obtain equidistant data and S13: data denoising smoothing;Step S2 includes substep S21: regression model establishment and S22: determine regression coefficient;Step S3 includes substep, S31: the determination method of limit stress static threshold of each monitoring point, S32: the determination method of stress change rate static threshold, S33: sharp end mutation model, the determination method of section stability threshold and S34: multi-factor joint early warning method.The application is by monitoring data preprocessing, data trend regression analysis and division early warning method and index. According to this, the overall trend and law of stress change of roadway surrounding rock can be accurately analyzed, and the anchor rod early warning system early warning threshold is accurately determined.

The invention discloses a track precision measurement method based on design file matching. The method comprises the steps of design file resolving, time-space synchronization of multiple sensors and segmented matching of design files, inertial navigation resolving initialization, inertial navigation resolving, indoor and outdoor scene discrimination, data fusion resolving and reverse smoothing. The problem that an existing scheme cannot be suitable for various detection scenes is solved; the method has the advantages that automatic identification and full coverage of the track detection scene can be achieved, the reliability is improved through a software and hardware combination mode for different scenes, design files and historical observation data are effectively integrated, the method is suitable for various detection scenes on the basis that the measurement precision is met, and establishment of a track detection database is facilitated. The invention further discloses a track precision measurement system based on design file matching and a track precision measurement device based on design file matching.

Embodiments disclose a method and a system for robust transformation of input with a neural network. The method comprises processing the input data with a variational neural network (VNN) trained with ML to produce static parameters including noise level for the input data, injecting a set of random noises sampled on a probabilistic distribution according to the statistic parameters defined by the VNN to produce a set of perturbed input samples. The method comprises processing each of the set of perturbed input samples with a transformation neural network to produce a set of transformations and outputting a combination of the set of transformations as the robust transformation of the input data. Some embodiments consider training the variational neural network and transformation neural network by using adversarial examples from an attack model via alternating, explicit, and implicit gradient frameworks.

The invention discloses a multi-scale time sequence modeling video abstract generation method fusing semantic enhancement and boundary perception, and belongs to the technical field of computer vision. The method comprises the following steps: carrying out feature coding on an input video frame sequence, and extracting a corresponding time sequence feature representation; inputting the extracted time sequence features into a video abstract generation model, and outputting importance scores by the model; performing time sequence smoothing and peak value detection on the importance score, and determining a key frame candidate set; and generating a final video abstract fragment based on the key frame candidate set in combination with a non-maximum suppression and fragment combination strategy. Experimental results obtained on reference data sets SumMe and TVSum prove the advancement of the method. According to the video abstract generation method provided by the invention, the problems of fuzzy event boundary and insufficient multi-scale time sequence understanding in the existing video abstract generation method can be effectively solved.

The invention belongs to the technical field of automated guided vehicles, and particularly relates to a multi-source positioning fusion and navigation correction method and system for an AGV, and the method comprises the steps: carrying out the pose estimation through the bidirectional coupling of extended Kalman filtering and particle filtering, and the estimation uncertainty is fed back to the extended Kalman filter to dynamically adjust the noise parameter. And further introducing sliding window factor graph optimization to carry out back-end smoothing. In the navigation stage, a dynamic decision maker based on deep reinforcement learning is adopted, and optimal control parameters are generated according to real-time positioning, map semantics and historical performance. The system can also enable the filtering model and the decision strategy to adapt to the current environment through periodic online fine tuning. Through innovatively and deeply fusing a classical state estimation method and a leading-edge machine learning method, the positioning precision, the environment understanding capability and the navigation intelligence of the AGV in a complex dynamic scene are remarkably improved.

The invention relates to the technical field of electric data processing, in particular to an air compressor operation data collecting method and system based on the Internet of Things, and the method comprises the steps that operation data of an air compressor are collected and preprocessed; any moment in any dimension is selected as a target moment, and a sequence of operation data of the target moment in the neighbor time window is acquired; and performing ascending sorting on the instantaneous energy of each moment in the sequence to obtain an instantaneous energy quantum sequence. According to the method, the standard deviation parameter in the Gaussian filtering algorithm is adaptively adjusted based on the global fault confidence, so that the method has relatively high smoothing capability on random noise data, the denoising capability is enhanced, more detailed information is reserved for fault feature data, excessive smoothing is avoided, and the fault detection accuracy is improved. Finally, data which can eliminate noise interference to the maximum extent and can retain real fault features are obtained, and high-quality data are provided for follow-up high-precision state monitoring and maintenance.

A navigation path can be determined for an object using one or more neural networks. In various embodiments, image data is obtained that is representative of an environment in which the object is to be navigated. Relevant features are identified from the image, and a curve fit to those features. Loss values for the potential paths are scaled based at least in part upon the distance of those features in the real world. This can include, in at least some embodiments, performing the scaling as a function of the curvature of the curve fit to the features. Temporal smoothing can be performed with respect to prior path predictions in order to prevent sudden changes in the predicted path. The paths are analyzed to select a path with a highest confidence value that also at least satisfies a minimum confidence criterion. The path can be converted into three-dimensional navigation information.