Region Enhanced Edge-Based Multi-Class Object Proposal for Self-Driving Vehicles

Abstract

On-road object detection is a fundamental element for the safety and reliability of autonomous driving systems. A primary challenge is developing object detection algorithms that are both fast and robust. This paper introduces a novel object proposal algorithm, named Region Enhanced Edge-Based (REEB) proposal, designed to accelerate object detection by significantly reducing the number of candidate regions requiring evaluation by a subsequent classification network. REEB leverages edge-map cues to score and rank initial proposals. To further enhance both detection quality and processing speed, the algorithm integrates efficient complementary techniques: image entropy is used to guide proposal generation density in relevant image regions, and road segmentation aids in refining proposal scores by differentiating road from non-road areas. Experimental evaluations on the KITTI dataset demonstrate that REEB achieves an average recall rate of 72.1% across four classes (pedestrian, cyclist, car, and truck) with an average processing time of 15 milliseconds per image. These results indicate strong performance when compared to other traditional, non-deep learning object proposal algorithms.