Refine
Document Type
- Article (1)
- Doctoral Thesis (1)
- Working Paper (1)
Language
- English (3)
Has Fulltext
- yes (3)
Keywords
- multi-object tracking (2)
- object classification (2)
- ADAS (1)
- autonomous driving (1)
- autonomous systems object classification (1)
- classification (1)
- clustering (1)
- environment perception (1)
- machine learning (1)
- radar (1)
Faculty / Organisational entity
Road accidents remain as one of the major causes of death and injuries globally. Several million people die every year due to road accidents all over the world. Although the number of accidents in European region have reduced in the past years, road safety still remains a major challenge. Especially in case of commercial trucks, due to the size and load of the vehicle, even minor collisions with other road users would lead to serious injuries or death. In order to reduce number of accidents, automotive industry is rapidly developing advanced driver assistance systems (ADAS) and automated driving technologies. Efficient and reliable solutions are required for these systems to sense, perceive and react to different environmental conditions. For vehicle safety applications such as collision avoidance with vulnerable road users (VRUs), it is not only important for the system to efficiently detect and track the objects in the vicinity of the vehicle but should also function robustly.
An environment perception solution for application in commercial truck safety systems and for future automated driving is developed in this work. Thereby a method for integrated tracking and classification of road users in the near vicinity of the vehicle is formulated. The drawbacks in conventional multi-object tracking algorithms with respect to state, measurement and data association uncertainties have been addressed with the recent advancements in the field of unified multi-object tracking solutions based on random finite sets (RFS). Gaussian mixture implementation of the recently developed labeled multi-Bernoulli (LMB) filter [RSD15] is used as the basis for multi-object tracking in this work. Measurement from an high-resolution radar sensor is used as the main input for detecting and tracking objects.
On one side, the focus of this work is on tracking VRUs in the near vicinity of the truck. As it is beneficial for most of the vehicle safety systems to also know the category that the object belongs to, the focus on the other side is also to classify the road users. All the radar detections believed to originate from a single object are clustered together with help of density based spatial clustering for application with noise (DBSCAN) algorithm. Each cluster of detections would have different properties based on the respective object characteristics. Sixteen distinct features based on radar detections, that are suitable for separating pedestrians, bicyclists and passenger car categories are selected and extracted for each of the cluster. A machine learning based classifier is constructed, trained and parameterised for distinguishing the road users based on the extracted features.
The class information derived from the radar detections can further be used by the tracking algorithm, to adapt the model parameters used for precisely predicting the object motion according to the category of the object. Multiple model labeled multi-Bernoulli filter (MMLMB) is used for modelling different object motions. Apart from the detection level, the estimated state of an object on the tracking level also provides information about the object class. Both these informations are fused using Dempster-Shafer theory (DST) of evidence, based on respective class probabilities Thereby, the output of the integrated tracking and classification with MMLMB filter are classified tracks that can be used by truck safety applications with better reliability.
The developed environment perception method is further implemented as a real-time prototypical system on a commercial truck. The performance of the tracking and classification approaches are evaluated with the help of simulation and multiple test scenarios. A comparison of the developed approaches to a conventional converted measurements Kalman filter with global nearest neighbour association (CMKF-GNN) shows significant advantages in the overall accuracy and performance.
This paper presents a method for classifying traffic participants based on high-resolution automotive radar sensors for autonomous driving applications. The major classes of traffic participants addressed in this work are pedestrians, bicyclists and passenger cars. The preprocessed radar detections are first segmented into distinct clusters using density-based spatial clustering of applications with noise (DBSCAN) algorithm. Each cluster of detections would typically have different properties based on the respective characteristics of the object that they originated from. Therefore, sixteen distinct features based on radar detections, that are suitable for separating pedestrians, bicyclists and passenger car categories are selected and extracted for each of the cluster. A support vector machine (SVM) classifier is constructed, trained and parametrised for distinguishing the road users based on the extracted features. Experiments are conducted to analyse the classification performance of the proposed method on real data.
This paper presents a method for simultaneous classification and robust tracking of traffic participants based on the labeled random finite set (RFS) tracking framework. Specifically, a method to integrate the object class information into the tracking loop of the multiple model labeled multi-Bernoulli (MMLMB) filter, using Dempster-Shafer evidence theory is presented. The multi-object state is estimated using the detections from the sensors and by propagation of multi-object density in a Bayesian fashion. Parallelly, the object class information is also predicted and updated recursively. The underlying object class information required for this could typically be obtained from different types of sensor such as radar, lidar and camera, using classical perception or more recent deep learning methods. On one hand, this enables an unified classification and tracking of traffic participants. On the other hand, it also increases the robustness of multi-object tracking, as the parameters of the tracking algorithm could be adapted using the class information. Moreover, using the Dempster-Shafer method for fusing class information from different sensor sources improves the overall performance, especially when the sensors have contradicting classification.