A Deep Unified Pedestrian Detection Framework

Pedestrian detection is one of the important applications of computer vision with key contributions in various fields of human life such as intelligent vehicles, surveillance, and advanced robotics. Considerable research has taken place in protecting vulnerable road users particularly pedestrians due to the high risk of accidents to ensure safety on roads. The process of pedestrian detection is often marred by significant intra-class variance due to varied postures and appearances, exhibited by humans on roads. Other crucial issues like occlusion, background complexity, lead to loss of accuracy in final detection results. Detecting overlapping, occluded, and small objects in a complex dataset like pedestrians, with acceptable accuracy, has always been a challenging task. The proposed work unifies two families of detectors, the first is one stage detector responsible for precise bounding boxes but low recall value while the other one is a family of two-stage detectors, giving a high recall value but an imprecise number of bounding boxes. Adaptive fusion of two generates enhanced detection accuracy and decreases the overall Log Average Miss Rate (LAMR). The performance of the proposed work has been evaluated and assessed on three publicly available datasets: ETH, INRIA, and Central Pedestrian crossing sequence, which exhibits superior pedestrian detection performance over the existing state-of-the-art.