Robust Pedestrian Detection and Path Prediction using Improved YOLOv5

Abstract

Pedestrian detection and path prediction are critical concerns in vision-based surveillance systems. Advanced computer vision applications are challenged by discrepancies in pedestrian postures, scales, backdrops, and occlusions. In the state-of-the-art approach, these challenges cannot be addressed because of limitations in postures and multiscale information missing in the features of the region of interest. In order to address these challenges, we present an improved YOLOv5-based deep learning approach. Occupied pedestrians are detected on multi-scales using the improved YOLOv5 model. We improved the YOLOv5 detection method in three ways: 1) a new feature fusion layer has been added to capture more shallow feature information of small size pedestrians, 2) features from the backbone network have been brought into the feature fusion layers to reduce feature information loss of small size pedestrians; and 3) Scale Invariant Cross-stage Partial Network (SCSP) has been added to detect the pose and scale invariant pedestrians. At last, the proposed path prediction method used to estimate the pedestrians path based on motion data. The proposed method deals with partial occlusion circumstances to reduce object occlusion-induced progression and loss, and links recognition results with motion attributes. It uses motion and directional data to estimate the movements and orientation of pedestrians. The proposed method outperforms the existing methods, according to the results of the experiments. Finally, we conclude and look into future research directions.