Information feedback loop for improved pedestrian detection in an autonomous perception system

Environmental perception systems are often built using technologies that operate in a sequential manner. In the task of tracking in particular, where the classical detector-tracker interaction is a serial process, it is viable to break this design rule by introducing information loops. This is especially feasible in a tracker that operates in a prediction-update cycle. Tracking predictions can steer object detection towards regions where an object is anticipated and, in turn, tracking updates can be improved by incorporating reinforced detections. In this paper we propose a novel detector-tracker feedback loop for information exchange based on spatio-temporal similarity of detections and tracklets. We reinforce pedestrian detections that have weak confidence scores by matching their bounding boxes to estimated tracklets with high tracking confidence. The proposed system has several compelling advantages: based on a positive feedback principle it extracts the maximum detection and tracking information, while operating transparently and with minimal computational load. In a controlled ablation study we evaluate our feedback mechanism using the KITTI object tracking dataset. We show that our system gains significant performance increase over the baseline in both frame-by-frame detection and tracking quality.