Surveillance Video Target Detection With Stitched Multi-Resolution Windows
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing surveillance edge computing systems face challenges in achieving high object detection accuracy, frame rate, and detection distance due to limited computing power and bandwidth, with current optimization methods being unsuitable for complex surveillance scenarios that require detection of both near and distant objects.
Innovation Solution
A method that utilizes the size and motion vector differences of objects in surveillance video to combine low frame rate detection of distant small objects with high frame rate detection of near objects, using a stitched image with varying resolutions and sliding windows to optimize detection distance and reduce computing power and bandwidth requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Length of stationary object
If multi-scale scaling is performed on the original input image to generate multi-scale pyramid image group for detecting objects at different distances, then detection distance is improved, but device complexity and computing power requirements increase significantly
Solution Approach 1:
The patent divides the detection task into two segments: full-image detection for distant objects and sliding window detection for near objects. This segmentation allows each part to use optimized processing - the full image is scaled down for distant object detection while sliding windows maintain higher resolution for near objects, avoiding the need to process the entire image at multiple scales
Solution Approach 2:
Different regions of the image are processed with different quality levels. Distant regions (captured in the full-image view) use lower resolution processing, while near regions (captured in sliding windows) use higher resolution processing. This local quality differentiation maintains detection accuracy where needed while reducing overall computing power requirements
2Productivity
If neural network algorithms with small input size are used to reduce computing power requirements, then productivity is improved, but measurement precision deteriorates due to reduced image resolution
Solution Approach 1:
The patent dynamically adjusts the input image size based on the detection region. For distant objects, a smaller scaled-down full image is used. For near objects, larger sliding window regions are extracted and processed at higher resolution. This dynamic adaptation allows the system to optimize computing power usage while maintaining detection accuracy for objects at different distances
Solution Approach 2:
The patent adds a spatial dimension to the processing by creating multiple views of the same scene at different scales and positions. Instead of processing one large image, it processes a full-image view plus multiple sliding window views, each contributing to detection accuracy in their respective regions
3Productivity
If key frame detection with low-complexity tracking is used to reduce computing power requirements, then productivity is improved, but reliability deteriorates when objects move fast due to missed detections
Solution Approach 1:
Instead of performing full detection on every frame, the patent performs full detection periodically on key frames and uses lighter tracking on intermediate frames. The sliding window mechanism ensures that even fast-moving objects are captured in subsequent key frames, providing a balance between frame rate and reliability
4Device complexity
If pruning and low-bit quantization are applied to simplify the neural network, then computing power requirements are reduced, but measurement precision deteriorates due to increased missed detection and false detection rates
Solution Approach 1:
The patent segments the detection process into full-image detection and sliding window detection, allowing the use of optimized neural network configurations for each segment without requiring aggressive pruning or quantization that would degrade overall accuracy
Data Source
AI summary
This application discloses a method and apparatus for improving video target detection performance in surveillance edge computing. This application relates to the technical field of digital image processing. The method includes: determine the size of multiple rectangular sliding windows for scanning according to the input size of the object detection neural network algorithm and the size of the original input image; when each frame is detected, the original input image and the sub-images in each rectangular sliding window are scaled in different proportions; the resolution of the scaled original input image is lower than that of the scaled sliding window sub-images; stitching the scaled images into a rectangular image and using it as a detection input image; the detection is performed by an object detection neural network algorithm corresponding to the size of the detection input image.