On the safety of vulnerable road users by cyclist orientation detection using Deep Learning

In this work, orientation detection using Deep Learning is acknowledged for aparticularly vulnerable class of road users,the cyclists. Knowing the cyclists'orientation is of great relevance since it provides a good notion about theirfuture trajectory, which is crucial to avoid accidents in the context ofintelligent transportation systems. Using Transfer Learning with pre-trainedmodels and TensorFlow, we present a performance comparison between the mainalgorithms reported in the literature for object detection,such as SSD, FasterR-CNN and R-FCN along with MobilenetV2, InceptionV2, ResNet50, ResNet101feature extractors. Moreover, we propose multi-class detection with eightdifferent classes according to orientations. To do so, we introduce a newdataset called "Detect-Bike", containing 20,229 cyclist instances over 11,103images, which has been labeled based on cyclist's orientation. Then, the sameDeep Learning methods used for detection are trained to determine the target'sheading. Our experimental results and vast evaluation showed satisfactoryperformance of all of the studied methods for the cyclists and theirorientation detection, especially using Faster R-CNN with ResNet50 proved to beprecise but significantly slower. Meanwhile, SSD using InceptionV2 providedgood trade-off between precision and execution time, and is to be preferred forreal-time embedded applications.