Enhancing video object Detection through feature extraction-based video compression

Projects* | 2023 | Links Demo | Source

In this project, we delve into the realm of video object detection to explore the fascinating intersection of video compression and object detection. Our primary objective is to enhance video object detection by using feature enhancing maps for video compression.

Problem Statement

Problem:

• Traditional video compression methods are tailored to human vision.
• Applications like automated video surveillance rely on computers to analyze the video frames.
• Traditional video compression often compromises object recognition precision.

Proposal

Enhancing video object detection by employing feature extraction-based video compression technique. Methodology:

1. Step 1: Extract features from RAW video frames.
2. Step 2: Calculate Object Importance maps from features.
3. Step 3: Devise Bit Allocation strategy from the Object Importance map.
4. Step 4: Use Bit allocation strategy to encode video frames.

Methodology

Step 1

• Pass the frames to a YOLO model, and extract features from CNN Layers of the model.
• Features essentially highlight important areas and offer insightful information of the frame.

Step 2

• CNN has 32 kernels in its first layer.
• Condense the features from each kernel by multiplying the average of each kernel.
• Calculate the L2 Normalized value for each kernel and average it.
• This is the object importance value for each pixel.

Step 3

• Divide the object importance maps into blocks.
• Calculate the QP values for each block.
• Store the QP values in a file so it can be used for video compression.

Step 4

• Pass the QP values into the FFMPEG video compressor

Results

The successful implementation of this project holds transformative potential in the medical field. Rapid and automated age prediction from chest X-rays can significantly enhance diagnostic workflows, aiding healthcare professionals in making informed decisions promptly. This project stands as a testament to the fusion of technology and healthcare for improved patient outcomes.

Conclusion

The proposed methodology showcases a promising approach to enhance object detection accuracy through customized video compression. The integration improved performance compared to traditional methods in terms of:

1. Higher confidence in predictions
2. Increased number of true detections
3. Decreased number of false detections
4. Better PSNR vs BitRate