Original Article Proc IMechE Part P: J Sports Engineering and Technology 2019, Vol. 233(4) 478–488 Ó IMechE 2019 Article reuse guidelines: sagepub.com/journals-permissions DOI: 10.1177/1754337119850927 journals.sagepub.com/home/pip Pattern recognition neural classifier for fall detection in rock climbing Angelo Bonfitto , Andrea Tonoli, Stefano Feraco , Enrico Cesare Zenerino and Renato Galluzzi Abstract From an athlete’s perspective, the identification of falls during rock climbing is of major importance. It constitutes a solid performance indicator, but more importantly, it could be used to trigger an instantaneous alarm to rescue teams, thus reducing the negative health consequences for the climber. In this context, an artificial neural network–based technique for fall detection during rock climbing is presented in this study. The output of this tool could be used for safety and per- formance monitoring purposes. The proposed method exploits a neural network for binary pattern recognition. This network is fed with a set of features extracted in real time from the acceleration and altitude signals acquired by means of a wearable device. The classifier is trained and validated with experimental datasets recorded during real climbing ses- sions of eight athletes through different route grades and conditions. This article illustrates the architecture of the pro- posed algorithm, feature extraction process, and evaluation of its accuracy. In addition, an analysis of the severity level of the detected falls is conducted. The method is able to identify real fall events with a high success rate, while yielding very few false positive indications of a fall. Keywords Artificial neural networks, classifier, pattern recognition, fall detection, rock climbing, acceleration, altitude, wearable device Date received: 19 December 2018; accepted: 20 April 2019 Introduction In the last decade, outdoor rock-climbing activities have witnessed a growing popularity worldwide. At the same time, wearable solutions to monitor the perfor- mance of athletes have shown a substantial diffusion. The availability of compact devices enables the concep- tion of real-time data logging and signal processing sys- tems, thus guaranteeing a high degree of portability, accuracy, and autonomy. 1–5 In particular, the potential safety risks of the activity motivate the development of algorithms for harmful event identification. Typically, these techniques are based on the direct measurement of sensed signals, such as pressure, altitude, and climber’s movements. Among the solutions dedicated to safety, the detection of high- risk falls is of major importance for outdoor climbing activities in remote environments. If accurate and reli- able, this task is crucial to trigger the fast response of rescue teams, hence reducing the negative consequences of the fall on the climber. In addition, the detection of the fall and the estimate of its severity can be useful to assess the safety chain, including rope, harness, carabi- ners, and others. Generally speaking, the problem of human falls is extensively addressed in the literature, but most of the reported studies focus on the cases of the elderly, 6–13 people with disabilities, 14 and personnel working in dangerous conditions, such as firefighters. 15 Although effective, these solutions cannot be adapted to the spe- cific context of outdoor rock climbing: they require indoor devices, such as cameras, 14–18 and they imply fall events with less severe dynamics. 19 As a matter of fact, rock-climbing activities are characterized by a wide range of dynamic contents, such as jumping over an obstacle, moving on rough terrain, or receiving shocks on the wearable sensor when rubbing it against a wall. These occurrences are characterized by high accelera- tion spikes and relevant altitude variations. However, Mechatronics Laboratory, Department of Mechanical and Aerospace Engineering, Politecnico di Torino, Torino, Italy Corresponding author: Angelo Bonfitto, Mechatronics Laboratory, Department of Mechanical and Aerospace Engineering, Politecnico di Torino, Corso Duca degli Abruzzi 24, I-10129 Torino, Italy. Email: angelo.bonfitto@polito.it