engleski

Improving fleet management systems by computer vision

This thesis proposes an image categorization framework to deliver added value to fleet management systems. In particular, this framework aims to improve map matching, route reconstruction, alarming and reporting. In order to match the client-server nature of fleet management the framework is conceived around the following two requirements: i) the bandwidth should be used sparingly, and ii) the set of image categories must be open. These requirements can be satisfied by a suitable division of responsibility between the clients and the server. The clients are responsible for representing images with descriptors which are designed to be compact and category-agnostic. The server is responsible for classifying descriptors into an arbitrary set of categories. This organization minimizes the bandwidth requirements due to compactness of the descriptors, and ensures that the set of categories remains open due to clients being oblivious to it. Several kinds of image descriptors have been considered: handcrafted gradient histograms (GIST, SIFT), spatial Fisher vector embeddings, and convolutional representations trained in an end-to-end fashion (VGG, DenseNet, ResNet, MobileNetV2 and DCGAN). The descriptors are further compressed using PCA and quantization, after which they are classified by SVM. In order to evaluate the considered methods we introduce FM3--a novel image dataset which is specifically designed for fleet management applications. The dataset contains 11448 images which were acquired in different weather conditions and labeled with the following binary attributes: highway, road, tunnel, tunnel exit, settlement, overpass, booth, traffic. The results indicate that excellent classification results can be achieved with deep convolutional representations trained in a supervised manner. We refrain from fine tuning on the target dataset (although this further improves the results) in order to avoid reducing the descriptor performance on new categories due to catastrophic forgetting. Image descriptors can be as small as 512 bits, while still offering good performance. The proposed framework is able to tolerate adverse weather and poor illumination conditions provided that some such samples are present in the SVM training dataset.

computer vision ; intelligent vehicles ; image classification

nije evidentirano