Human shape modelling for carried object detection and segmentation
|Advisor:||Bergevin, Robert; Bilodeau, Guillaume-Alexandre|
|Abstract:||Detecting carried objects is one of the requirements for developing systems that reason about activities involving people and objects. This thesis presents novel methods to detect and segment carried objects in surveillance videos. The contributions are divided into three main chapters. In the first, we introduce our carried object detector which allows to detect a generic class of objects. We formulate carried object detection in terms of a contour classification problem. We classify moving object contours into two classes: carried object and person. A probability mask for person’s contours is generated based on an ensemble of contour exemplars (ECE) of walking/standing humans in different viewing directions. Contours that are not falling in the generated hypothesis mask are considered as candidates for carried object contours. Then, a region is assigned to each carried object candidate contour using Biased Normalized Cut (BNC) with a probability obtained by a weighted function of its overlap with the person’s contour hypothesis mask and segmented foreground. Finally, carried objects are detected by applying a Non-Maximum Suppression (NMS) method which eliminates the low score carried object candidates. The second contribution presents an approach to detect carried objects with an innovative method for extracting features from foreground regions based on their local contours and superpixel information. Initially, a moving object in a video frame is segmented into multi-scale superpixels. Then human-like regions in the foreground area are identified by matching a set of extracted features from superpixels against a codebook of local shapes. Here the definition of human like regions is equivalent to a person’s probability map in our first proposed method (ECE). Our second carried object detector benefits from the novel feature descriptor to produce a more accurate probability map. Complement of the matching probabilities of superpixels to human-like regions in the foreground are considered as a carried object probability map. At the end, each group of neighboring superpixels with a high carried object probability which has strong edge support is merged to form a carried object. Finally, in the third contribution we present a method to detect and segment carried objects. The proposed method adopts the new superpixel-based descriptor to identify carried object-like candidate regions using human shape modeling. Using spatio-temporal information of the candidate regions, consistency of recurring carried object candidates viewed over time is obtained and serves to detect carried objects. Last, the detected carried object regions are refined by integrating information of their appearances and their locations over time with a spatio-temporal extension of GrabCut. This final stage is used to accurately segment carried objects in frames. Our methods are fully automatic, and make minimal assumptions about a person, carried objects and videos. We evaluate the aforementioned methods using two available datasets PETS 2006 and i-Lids AVSS. We compare our detector and segmentation methods against a state-of-the-art detector. Experimental evaluation on the two datasets demonstrates that both our carried object detection and segmentation methods significantly outperform competing algorithms.|
|Document Type:||Thèse de doctorat|
|Open Access Date:||31 August 2018|
|Collection:||Thèses et mémoires|
All documents in CorpusUL are protected by Copyright Act of Canada.