Cyclist detection in LIDAR scans using faster R-CNN and synthetic depth images

In this paper a vision-based system for cyclist detection for highly automated vehicles is presented which utilize a deep convolutional neural network. Given the limited amount of LiDAR data for cyclists for the task of cyclist detection, a realistic synthetic data generation pipeline has been adopted in this work to overcome this problem by generate virtually unlimited number of synthetic training, testing and validation depth image datasets of cyclists. Then, using a holistic detection technique, any instances of cyclists can be localised in the generated synthetic depth images which can be shown as a down-sampled instances of 3D LiDAR data. The proposed technique in this paper, exploits the Faster Region-based Convolutional Network (Faster R-CNN) architecture for simultaneously detecting and localising any instances of cyclists in depth images. The trained Faster R-CNN models have been tested on the widely popular KITTI urban dataset and the results show that the proposed framework outperforms the classical HOG+SVM object localisation method with increased 21% in average precision. This shows the ability of the trained models using the proposed framework to generalise from synthetic training dataset to operate on live dataset.