FSSPD: Fast Single Stage Pedestrian Detector for Autonomous Driving

For low accuracy of the anchor-based pedestrian detectors in the case of small and high-density pedestrian, a Fast Single Stage Pedestrian Detector (FSSPD) is presented, which has three merits. Firstly, a single stage anchor-based neural network is designed on the base of the modified Darknet-19 instead of the widely used VGG-16, largely reducing inference time during pedestrian detection. Secondly, Darknet-19 usually adapted in YOLOv2 is modified by using a scale invariant network structure with multi detection modules to detect small pedestrians. Thirdly, a dense anchor strategy is proposed to deal with high mistake rate and miss rate of high-density pedestrian. As a result, the proposed network structure and strategies are proved effective for small and highdensity pedestrian detection. On the evaluation set of KITTI dataset, dense anchor strategy improves the average precision by 1.5%, 1.5% and 1% in easy, moderate and hard level, respectively. And the multidetection- module strategy works well for small pedestrian with significantly improving the average precision by 15.8%, 16.5% and 17% in these three difficulty levels, respectively. On the test set of KITTI dataset, our approach get a comparative result compared to previous methods at the speed of 14.3fps in near real time. This indicates that FSSDP is suitable for pedestrian detection as a lightweight framework with a good tradeoff between precision and speed.