This letter considers the problem of detecting and tracking objects in a sequence of images. The problem is formulated in a filtering framework, using the output of object-detection algorithms as measurements. An extension to the filtering formulation is proposed that incorporates class information from the previous frame to robustify the classification. Further, the properties of the object-detection algorithm are exploited to quantify the uncertainty of the bounding box detection in each frame. The complete filtering method is evaluated on camera trap images of the four large Swedish carnivores, bear, lynx, wolf, and wolverine. The experiments show that the class tracking formulation leads to a more robust classification.

lassifiers based on neural networks (nn) often lack a measure of uncertainty in the predicted class. We propose a method to estimate the probability mass function (pmf) of the different classes, as well as the covariance of the estimated pmf. First, a local linear approach is used during the training phase to recursively compute the covariance of the parameters in the nn. Secondly, in the classification phase, another local linear approach is used to propagate the covariance of the learned nn parameters to the uncertainty in the output of the last layer of the nn. This allows for an efficient Monte Carlo (mc) approach for; (i) estimating the pmf; (ii) calculating the covariance of the estimated pmf; and (iii) proper risk assessment and fusion of multiple classifiers. Two classical image classification tasks, i.e., mnist, and cfar10, are used to demonstrate the efficiency of the proposed method.

Suppose data-driven black-box models, e.g., neural networks, should be used as components in safety-critical systems such as autonomous vehicles. In that case, knowing how uncertain they are in their predictions is crucial. However, this needs to be provided for standard formulations of neural networks. Hence, this thesis aims to develop a method that can, out-of-the-box, extend the standard formulations to include uncertainty in the prediction. The proposed method in the thesis is based on a local linear approximation, using a two-step linearization to quantify the uncertainty in the prediction from the neural network. First, the posterior distribution of the neural network parameters is approximated using a Gaussian distribution. The mean of the distribution is at the maximum a posteriori estimate of the parameters, and the covariance is estimated using the shape of the likelihood function in the vicinity of the estimated parameters. The second linearization is used to propagate the uncertainty in the parameters to uncertainty in the model’s output. Hence, to create a linear approximation of the nonlinear model that a neural network is. 

The first part of the thesis considers regression problems with examples of road-friction experiments using simulated and experimentally collected data. For the model-order selection problem, it is shown that the method does not under-estimate the uncertainty in the prediction of overparametrized models. 

The second part of the thesis considers classification problems. The concept of calibration of the uncertainty, i.e., how reliable the uncertainty is and how close it resembles the true uncertainty, is considered. The proposed method is shown to create calibrated estimates of the uncertainty, evaluated on classical image data sets. From a computational perspective, the thesis proposes a recursive update of the parameter covariance, enhancing the method’s viability. Furthermore, it shows how quantified uncertainty can improve the robustness of a decision process by formulating an information fusion scheme that includes both temporal correlational and correlation between classifiers. Moreover, having access to a measure of uncertainty in the prediction is essential when detecting outliers in the data, i.e., examples that the neural network has yet to see during the training. On this task, the proposed method shows promising results. Finally, the thesis proposes an extension that enables a multimodal representation of the uncertainty. 

The third part of the thesis considers the tracking of objects in image sequences, where the object is detected using standard neural network-based object detection algorithms. It formulates the problem as a filtering problem with the prediction of the class and the position of the object viewed as the measurements. The filtering formulation improves robustness towards false classifications when evaluating the method on examples from animal conservation in the Swedish forests. 

The uncertainty in the prediction calculated using the delta method for an over-parameterized (parametric) black-box model is shown to be larger or equal to the uncertainty in the prediction of a canonical (minimal) model. Equality holds if the additional parameters of the overparameterized model do not add flexibility to the model. As a conclusion, for an overparameterized black-box model, the calculated uncertainty in the prediction by the delta method is not underestimated. The results are shown analytically and are validated in a simulation experiment where the relationship between the normalized traction force and the wheel slip of a car is modelled using e.g., a neural network.

Neural Networks (NNS) can solve very hard classification and estimation tasks but are less well suited to solve complex sensor fusion challenges, such as end-to-end control of autonomous vehicles. Nevertheless, NN can still be a powerful tool for particular sub-problems in sensor fusion. This would require a reliable and quantifiable measure of the stochastic uncertainty in the predictions that can be compared to classical sensor measurements. However, current NN'S output some figure of merit, that is only a relative model fit and not a stochastic uncertainty. We propose to embed the NN'S in a proper stochastic system identification framework. In the training phase, the stochastic uncertainty of the parameters in the (last layers of the) NN is quantified. We show that this can be done recursively with very few extra computations. In the classification phase, Monte-Carlo (MC) samples are used to generate a set of classifier outputs. From this set, a distribution of the classifier output is obtained, which represents a proper description of the stochastic uncertainty of the predictions. We also show how to use the calculated uncertainty for outlier detection by including an artificial outlier class. In this way, the NN fits a sensor fusion framework much better. We evaluate the approach on images of handwritten digits. The proposed method is shown to be on par with MC dropout, while having lower computational complexity, and the outlier detection almost completely eliminates false classifications.

Despite the great success of neural networks (NN) in many application areas, it is still not obvious how to integrate an NN in a sensor fusion framework. The reason is that the computation of the for fusion required variance of NN is still a rather immature area. Here, we apply a methodology from system identification where uncertainty of the parameters in the NN are first estimated in the training phase, and then this uncertainty is propagated to the output in the prediction phase. This local approach is based on linearization, and it implicitly assumes a good signal-to-noise ratio and persistency of excitation. We illustrate the proposed method on a fundamental problem in advanced driver assistance systems (ADAS), namely to estimate the tire-road friction. This is a single input single output static nonlinear relation that is simple enough to provide insight and it enables comparisons with other parametric approaches. We compare both to existing methods for how to assess uncertainty in NN and standard methods for this problem, and evaluate on real data. The goal is not to improve on simpler methods for this particular application, but rather to validate that our method is on par with simpler model structures, where output variance is immediately provided.

In science, technology, and engineering, creating models of the environment to predict future events has always been a key component. The models could be everything from how the friction of a tire depends on the wheels slip  to how a pathogen is spread throughout society.  As more data becomes available, the use of data-driven black-box models becomes more attractive. In many areas they have shown promising results, but for them to be used widespread in safety-critical applications such as autonomous driving some notion of uncertainty in the prediction is required.

An example of such a black-box model is neural networks (NNs). This thesis aims to increase the usefulness of NNs by presenting an method where uncertainty in the prediction is obtained by linearization of the model. In system identification and sensor fusion, under the condition that the model structure is identifiable, this is a commonly used approach to get uncertainty in the prediction from a nonlinear model. If the model structure is not identifiable, such as for NNs, the ambiguities that cause this have to be taken care of in order to make the approach applicable. This is handled in the first part of the thesis where NNs are analyzed from a system identification perspective, and sources of uncertainty are discussed.

Another problem with data-driven black-box models is that it is difficult to know how flexible the model needs to be in order to correctly model the true system. One solution to this problem is to use a model that is more flexible than necessary to make sure that the model is flexible enough. But how would that extra flexibility affect the uncertainty in the prediction? This is handled in the later part of the thesis where it is shown that the uncertainty in the prediction is bounded from below by the uncertainty in the prediction of the model with lowest flexibility required for representing true system accurately. 

In the literature, many other approaches to handle the uncertainty in predictions by NNs have been suggested, of which some are summarized in this work. Furthermore, a simulation and an experimental studies inspired by autonomous driving are conducted. In the simulation study, different sources of uncertainty are investigated, as well as how large the uncertainty in the predictions by NNs are in areas without training data. In the experimental study, the uncertainty in predictions done by different models are investigated. The results show that, compared to existing methods, the linearization method produces similar results for the uncertainty in predictions by NNs.

An introduction video is available at https://youtu.be/O4ZcUTGXFN0

Inom forskning och utveckling har det har alltid varit centralt att skapa modeller av verkligheten. Dessa modeller har bland annat använts till att förutspå framtida händelser eller för att styra ett system till att bete sig som man önskar. Modellerna kan beskriva allt från hur friktionen hos ett bildäck påverkas av hur mycket hjulen glider till hur ett virus kan sprida sig i ett samhälle. I takt med att mer och mer data blir tillgänglig ökar potentialen för datadrivna black-box modeller. Dessa modeller är universella approximationer vilka ska kunna representera vilken godtycklig funktion som helst. Användningen av dessa modeller har haft stor framgång inom många områden men för att verkligen kunna etablera sig inom säkerhetskritiska områden såsom självkörande farkoster behövs en förståelse för osäkerhet i prediktionen från modellen.

Neuronnät är ett exempel på en sådan black-box modell. I denna avhandling kommer olika sätt att tillförskaffa sig kunskap om osäkerhet i prediktionen av neuronnät undersökas. En metod som bygger på linjärisering av modellen för att tillförskaffa sig osäkerhet i prediktionen av neuronnätet kommer att presenteras. Denna metod är välbeprövad inom systemidentifiering och sensorfusion under antagandet att modellen är identifierbar. För modeller såsom neuronnät, vilka inte är identifierbara behövs det att det tas hänsyn till tvetydigheterna i modellen.

En annan utmaning med datadrivna black-box modeller, är att veta om den valda modellmängden är tillräckligt generell för att kunna modellera det sanna systemet. En lösning på detta problem är att använda modeller som har mer flexibilitet än vad som behövs, det vill säga en överparameteriserad modell.  Men hur påverkas osäkerheten i prediktionen av detta? Detta är något som undersöks i denna avhandling, vilken visar att osäkerheten i den överparameteriserad modellen kommer att vara begränsad underifrån av modellen med minst flexibilitet som ändå är tillräckligt generell för att modellera det sanna systemet. Som avslutning kommer dessa resultat att demonstreras i både en simuleringsstudie och en experimentstudie inspirerad av självkörande farkoster. Fokuset i simuleringsstudien är hur osäkerheten hos modellen är i områden med och utan tillgång till träningsdata medan experimentstudien fokuserar på jämförelsen mellan osäkerheten i olika typer av modeller.Resultaten från dessa studier visar att metoden som bygger på linjärisering ger liknande resultat för skattningen av osäkerheten i prediktionen av neuronnät, jämfört med existerande metoder.

The prediction uncertainty of a neural network is considered from a classical system identification point of view. To know this uncertainty is extremely important when using a network in decision and feedback applications. The asymptotic covariance of the internal parameters in the network due to noise in the observed dependent variables (output) and model class mismatch, i.e., the true system cannot be exactly described by the model class, is first surveyed. This is then applied to the prediction step of the network to get a closed form expression for the asymptotic, in training data information, prediction variance. Another interpretation of this expression is as the non-asymptotic Cramér-Rao Lower Bound. To approximate this expression, only the gradients and residuals, already computed in the gradient descent algorithms commonly used to train neural networks, are needed. Using a toy example, it is illustrated how the uncertainty in the output of a neural network can be estimated.