Abstract

Artificial intelligence-based object detection models play a vital role in recognizing pedestrian and sending collision warnings in Advanced Driver Assistance Systems (ADAS). Although the deep learning-based modern detection algorithms provide accurate pedestrian detection under benchmark environment conditions, it has been shown by the research on unreliable pedestrian detection that detection results can be temporally unstable and delayed accumulation of confidence in a realistic scene. The delay or unstable behavior of a detection model in safety-critical applications such as driver assistance systems can cause problems in timely alert and response. In this paper, a system-level framework to achieve timeliness and reliable alerting without changing the ADAS object detection architecture is introduced. By using temporal confidence aggregation, detection continuity checking, and risk-based alerting, the proposed system utilizes the continuity of detections rather than solely spatial accuracy. Also, time-to-collision (TTC) is used to trigger an alert instead of detection of pedestrian only. It is verified by experiments in an urban pedestrian video datasets that this system has reliable warning behavior and oscillatory alerting is also alleviated without the loss of accuracy compared to a benchmark system. This work shows the importance of temporal consistency in AI safety systems and a potential way to achieve timely alerting in real ADAS.

Authors

G. Ananthi, G. Maha Vignesh, K.P. Balajiraj
Mepco Schlenk Engineering College, India

Keywords

Advanced Driver Assistance Systems (ADAS), Pedestrian Detection, Alert Timeliness, Time- to-Collision (TTC), Driver Safety