Improving Object Detection in Videos: A Comprehensive Evaluation of Faster R-CNN Employed in Partial Occlusion Handling
Jun Jie Yap, Yoke Leng Yong, Muhammed Basheer Jasser, Samuel-Soma M. Ajibade, Ismail Ahmed Al-Qasem Al-Hadi · 2024
Most occlusion handling research aims to tackle partial occlusion on images without the utilization of transferable contextual knowledge. To imitate the way a human performs solid object detection by utilizing contextual knowledge, a customized detection methodology utilising Faster Region-Based Convolutional Neural Network (Faster RCNN) is proposed. Building upon the original Faster R-CNN, our approach incorporates an auxiliary sub-model, complemented by OC-SORT for unique object tracking. Through continuous learning on the input frames, the sub-model acquires sufficient contextual knowledge for occlusion handling in upcoming frames, thereby enhancing the overall performance. The methodology is meticulously examined through comprehensive analyses, comparing the performance of the original Faster R-CNN implementation and the proposed methodology in several ways. An eight percent increase in the frequency of objects detected with L1 occlusion (0-20 percent occlusion) as well as a 93.66 per cent increment in the frequency of objects detected with L2 occlusion (20-40 percent occlusion) has been demonstrated, surpassing the baseline Faster R-CNN implementations.