Detection of Risky Riding Patterns of Motorcyclists based on Deep Learning and Linear Regression

Motorcycle accidents are the most fatal road accidents found in many regions especially in Asian countries. From the accident statistics, a major cause is due to riders’ riding behaviors such as fast, drunk or reckless ridings which are generally defined as abnormal riding. Detection of such riding pattern is challenging and would be beneficial for preventing possible road accidents. This paper proposes a novel framework to detect abnormal riding in three risky cases: weaving, swerving and drifting from recorded video footages. The methodology comprises of two main steps. First we localized motorcycles in video frames using Convolution Neural Network with a model namely ‘rfcn_resnet101_coco’. Second all detected centroids of motorcycles were fitted with two linear regression models i.e. Ordinary least square (OLS) and Random sample consensus (RANSAC) to find their linearity. The riding patterns whose regression scores are high tends to be normal ridings. From experiments, OLS and RANSAC showed a good performance to differentiate between normal and abnormal driving. The thresholds around 0.95 for OLS score or R squared and 0.94 for RANSAC score are sufficient for this classification. In addition, RANSAC provided an advantage over OLS when there exist noises e.g. nearby parking motorcycles.