Falls are a prominent cause of mortality and severe injury among the elderly. It can be prevented by tracking them and providing prompt treatment. Current fall detection systems use data from sensors or cameras in various ways. False positives are common in sensor-based systems, and operating system constraints make privacy a major concern in vision-based systems. This paper proposes a technique for detecting falls from RGB images using a convolutional neural network (CNN) utilizing movement trace characteristics generated by a modified structural similarity index (SSIM) that can be integrated into resource-constrained devices for in-house monitoring. The proposed approach uses a camera system and is tested against the UR Fall detection (URFD) dataset, outperforming previous fall detection systems. Our method achieves 99% accuracy. The model's dependence on readily available sensors and superior performance on the URFD dataset makes it a viable option for reliable fall detection in the real world.