Dogs housed in shelters may experience poor welfare. To ensure these dogs a good quality of life, welfare assessment tools should be sensitive not only to the animals’ physical health but also to their mental state, including the assessment of positive and negative emotions. In this study, we focused on the assessment of shelter dogs’ emotional expression using a Qualitative Behavioural Assessment (QBA) approach. Previous work successfully applied QBA to assess the emotional state of working and rescue dogs, and the observations were carried out on individual dogs in standardised settings with little or no stimulation. Results from such experiments might not be fully representative of the expressive demeanour that a dog could show in shelter conditions, where animals are exposed to a number of social and environmental stimuli. Thus, our aim was to apply QBA to a wider variety of shelter environments and social contexts than has been done so far, giving the animals the opportunity to express a wider repertoire of emotions and allowing for a more comprehensive assessment of dogs’ affective state. A set of descriptive terms was generated using Free-Choice-Profiling methodology by a group of 13 observers. QBA was made by scoring 16 video clips of shelter dogs in very different contexts (e.g. single/pair/group housing, presence/absence of human activity). Generalised Procrustes Analysis showed a high consensus between observers’ scoring patterns (75.7%; p < 0.001), and generated three main consensus dimensions explaining overall 66.6% of the variation between clips. The terms generated by the observers describing these consensus dimensions were semantically consistent, and characterised dogs as ranging: 1) from “playful/sociable/curious” to “bored/uncomfortable/apathetic”, 2) from “relaxed/tranquil” to “nervous/alert/fearful” and 3) from “stressed/anxious” to “wary/timorous/hesitant”. Overall, these broad dimensions are similar to those described in previous QBA studies on dogs. However, we detected differences in the type or frequency of the terms used, especially concerning three semantic spheres (i.e. “sociability”, “fearfulness” and “boredom”). It appears that, compared to what has been reported previously, by presenting more complex contexts and thus giving the animals the opportunity to express different behaviours, we generated a richer list of terms representing a wider repertoire of emotions. Our results support the notion that QBA can be immediately sensitive to an animal’s circumstances, integrating the ways in which animals experience the conditions in which they live into meaningful emotional indicators. This also highlights the importance of developing QBA tools that are species- and context-specific, especially for applied purposes.