Current whip violation detection in horse racing relies primarily on manual video reviews, an inefficient and labor-intensive process unsuitable for real-time enforcement. Given the growing concerns about animal welfare and fairness in racing, there is an urgent need to develop efficient and reliable automated methods to monitor whip usage. To address this, the present study specifically focuses on acoustically detecting whip strike sounds. However, detecting these whip sounds poses significant challenges due to their extremely short duration and variable acoustic characteristics, compounded by environmental noise such as crowd reactions, hoof impacts, and variable weather conditions. We utilize Convolutional Recurrent Neural Networks to tackle these complexities, initially achieving an F1-score of 69.8% with audio recorded from a single location using stereo microphones. Despite this promising result, obtaining high-quality acoustic data from horse races involves substantial logistical restrictions, including limited sensor placement options and significant sound attenuation over distance, limiting the effectiveness of single-location recordings. To overcome these limitations, we further integrate strategically placed microphone arrays combined with an ensemble detection approach. This allows capturing spatially diverse acoustic signals, effectively mitigating the constraints of single-location data acquisition and enhancing overall detection reliability.
Taguchi et al. (2025) studied this question.