Does the tail show when the nose knows? Artificial intelligence outperforms human experts at predicting detection dogs finding their target through tail kinematics

Abstract

Detection dogs are utilized for searching and alerting to various substances due to their olfactory abilities. Dog trainers report being able to 'predict' such identification based on subtle behavioural changes, such as tail movement. This study investigated tail kinematic patterns of dogs during a detection task, using computer vision to detect tail movement. Eight dogs searched for a target odour on a search wall, alerting to its presence by standing still. Dogs' detection accuracy against a distractor odour was 100% with trained concentration, while during threshold assessment, it progressively reached 50%. In the target odour area, dogs exhibited a higher left-sided tail-wagging amplitude. An artificial intelligence (AI) model showed a 77% accuracy score in the classification, and, in line with the dogs' performance, progressively decreased at lower odour concentrations. Additionally, we compared the performance of an AI classification model to that of 190 detection dog handlers in determining when a dog was in the vicinity of a target odour. The AI model outperformed dog professionals, correctly classifying 66% against 46% of videos. These findings indicate the potential of AI-enhanced techniques to reveal new insights into dogs' behavioural repertoire during odour discrimination.

Keywords: animal behaviour; artificial intelligence; computer vision; detection dogs; dog olfaction; domestic dog; visual communication.

Figure 1.

A schematic view of the study’s design. Step 1 (dog testing): dogs were previously trained and then tested on odour detection thresholds (test 1: performed on the trained concentration; test 2: decreasing odour concentrations until reaching detection threshold) with overhead video recording. Step 2 (manual video coding): manual video annotation. Step 3 (AI model 1—landmark detection): training landmarks detection of key dogs’ body from the videos. Step 4 (tail kinematics analysis): statistical analysis of tail kinematic features. Step 5 (AI model 2—time-series classifier): time-series data classification. Step 6 (comparison with experts): detection dog experts’ classifications on selected clips were compared to AI model accuracy.

Figure 2.

Image displaying the apparatus, testing room and three searching areas.

Figure 3.

(a) Schematic representation of a dog with a coordinate system and tail-wagging angle θ. (b) An example of the dynamics of an angle (left—negative angle) and angular velocity (right—positive angle) of the tail middle landmark during a random trial.

Figure 4.

Kinematics features profiles during test 1. Across panels, large markers depict group means and smaller markers individuals for non-target (left) and target areas (right). *p < 0.05.

Figure 5.

Dogs’ accuracy in signalling the target odour and the AI model’s accuracy in predicting whether dogs were in the target odour area or not, in test 1. Three area types were present: target, distractor and no odour. Chance level is indicated with the dotted line at 33%.

Figure 6.

Dogs’ accuracy in signalling the target odour, and the AI model’s accuracy in classifying whether dogs were in the target versus distractor odour areas in test 2 (with decreasing concentrations). Odour concentration 10⁻³ dilution factor was used for training and is not shown here. Three areas were present, but one contained the target odour and the other two contained the distractor odour, thus chance level is indicated with the dotted line at 50%.

Figure 7.

Accuracy in classifying whether dogs were in the target odour area or not by professional experts and the AI model.