Semantically Enhanced Multi-scale Feature Pyramid Fusion for Pedestrian Detection

Detecting multi-scale pedestrians (especially small scale ones) is one of the most challenging problems in computer vision community. At present, most existing pedestrian detectors only adopt single-scale feature map in their backbone network for detection, which is not capable of fully taking advantages of multi-scale feature information, and thus resulting in unsatisfactory multi-scale detection performance. To address this issue, we propose in this paper a semantically enhanced multi-scale feature pyramid fusion method that can effectively extract and integrate multi-scale feature maps for multi-scale pedestrian detection. The proposed method consists of two main components: 1) Trapezoidal Path Augmented Module (TPAM) and 2) Multi-scale Feature Fusion Module (MFFM). TPAM aims at extracting higher-level semantic features by the additional higher-level feature layers, where the produced features are enhanced with supplementary higher-level semantic information, so that they can focus more accurately in the pedestrian area, leading to improved detection performance. MFFM aims at integrating multi-scale feature maps coming from TPAM to further take advantages of multi-scale feature information and reduce computational redundancy caused by multiple detection heads. By extensive experimental evaluations on the popular CityPersons and Caltech benchmarks, our proposed method achieves superior performances than previous state of the arts on multi-scale pedestrian detection.