BGaitR-Net: An effective neural model for occlusion reconstruction in gait sequences by exploiting the key pose information

Somnath Sendhil Kumar, Binit Singh, Pratik Chattopadhyay, Agrya Halder, Lipo Wang

Abstract: Highlights•Introduces a new neural model BGaitR-Net to reconstruct occlusion in gait sequences.•CVAE and Bi-LSTM stacked within BGaitR-Net for embedding and sequence generation.•Auxiliary one-hot key pose vector guides BGaitR-Net to make improved frame prediction.•Dice score 0.98 and accuracy 95% obtained for 50% synthetic occlusion in CASIA-B data.•Outperforms other video prediction and occlusion handling methods in gait recognition.