Introduction
Fall detection is a critical technology for elderly care, yet it faces a significant "Goldilocks" problem. Traditional computer vision methods often rely on 3D networks to capture temporal movement, which are computationally expensive and memory-intensive, making them unsuitable for mobile or low-power embedded systems. Conversely, simpler 2D convolutional neural networks (CNNs) frequently struggle to distinguish the dynamic act of falling from the static state of lying down, as a single frame of a person on the floor can appear identical in both scenarios.
This research proposes a solution that is both lightweight and context-aware, bridging the gap between computational efficiency and semantic accuracy by combining motion analysis with a highly optimized neural network.
Methodology
Two-Stream Architecture
The proposed system utilizes a two-stream approach to balance performance and efficiency. As shown in Figure 1, the system processes video frames through a motion-triggered pipeline.
- Motion Characteristic Stream: This stream acts as a pre-filter or "gatekeeper." It employs a frame difference method to calculate the horizontal and vertical velocities of a subject. The second, more intensive stream is only triggered if the velocity exceeds a specific threshold, indicating a sudden, involuntary change in position. This effectively filters out static poses like lying down that lack the preceding high-velocity motion of a fall.
- MobileVGG Stream: Once triggered, the core classifier analyzes the scene. The authors optimized the classic VGG16 architecture into "MobileVGG" to minimize its memory footprint while maintaining high accuracy.
MobileVGG Optimization
MobileVGG achieves its efficiency through several architectural innovations designed to reduce the parameter count (the $R$ ratio) without losing VGG-level accuracy:
- Depthwise Separable Convolutions (DSC): Standard convolutions are replaced with a three-layer bottleneck structure: a pointwise convolution ($1 \times 1$), followed by a depthwise convolution ($3 \times 3$), and another pointwise convolution ($1 \times 1$). This drastically reduces computational cost.
- Residual Connections: To facilitate training in the 42-layer MobileVGG model and prevent vanishing gradients, the authors integrated skip connections, as illustrated in Figure 2.
- Global Average Pooling: Memory-intensive fully connected layers are replaced with $1 \times 1$ convolutions and global average pooling.
Experimental Setup and Results
The model was evaluated on a self-collected dataset of 3,628 images categorized into walking (1,176), falling on chairs (1,272), and falling to the ground (1,180). The dataset was captured using a webcam mounted 1.6 meters above the ground and split 1:1 for training and testing.
Performance Metrics
The results demonstrate a massive improvement in efficiency with minimal performance loss:
- Memory Footprint: For a $224 \times 224$ input, MobileVGG required only 24.7 MB, compared to 1555.7 MB for VGG16—a 63x reduction.
- Speed: Training for $100 \times 100$ inputs was significantly faster, taking 3.69s compared to 6.53s for VGG16.
- Accuracy: MobileVGG achieved a test accuracy of 98.25%, remarkably close to the performance of the much larger pre-trained VGG16.
Conclusion
By combining a computationally "cheap" motion gatekeeper with a highly optimized MobileVGG classifier, this approach enables high-accuracy fall detection on resource-constrained devices. The use of residual connections and depthwise separable convolutions allows the model to maintain deep feature representation while significantly reducing the parameter count. Future work will explore incorporating lightweight 3D network elements to further enhance temporal understanding and testing in more diverse, cluttered environments.
Summary updated