Multiple objects tracking by a highly decisive three-frame differencing-combined-background subtraction method with GMPFM-GMPHD filters and VGG16-LSTM classifier

摘要

Tracking of moving vehicles and pedestrians is the most important application in traffic surveillance videos. This study develops a highly efficient and fast multi-object tracking method using three-frame differencing-combined-background subtraction (TFDCBS)-coupled-automatic and fast histogram-entropy-based thresholding (HEBT) method together with GMPFM-GMPHD filters and VGG16-LSTM classifier. Here TFDCBS-HEBT methods identify the targeted objects with enclosed 3D bounding boxes and extracts multiple features from the raw images. Maximum number of error-free extracted multiple features (key points, multiple local convolutions, corners, and descriptors) are processed subsequently for object tracking by GMPFM-GMPHD Filters and an upgraded VGG16- LSTM classifier. The proposed method has been validated on KITTI 3D bounding boxdataset and its performance compared with three state-of-the-art tracking methods. Highest values of several performance parameters and the lowest computation time clearly demonstrate the promising feature of our new method for its application towards a fast and effective multi-target tracking of moving objects.