Vehicle Reidentification using YOLOv8 and feature-based similarity
Check .
In this project, we present a comprehensive vehicle re-identification system designed to enhance accuracy and robustness in surveillance footage analysis. Leveraging state-of-the-art object detection capabilities with YOLOv8, our system first identifies vehicles within frames. Subsequently, we employ a diverse array of feature extraction techniques that users can select, including Histogram of Oriented Gradients (HOG), color histograms, deep neural network (DNN) features via ResNet and OSNet, and Scale-Invariant Feature Transform (SIFT) descriptors. This diversity allows us to capture a broad spectrum of vehicle characteristics. Additionally, the system integrates Automatic Number Plate Recognition (ANPR) using PaddleOCR to enhance identification precision by verifying vehicle license plates. These features are then utilized for similarity comparison across frames, facilitating the accurate reidentification of vehicles over time. Through rigorous experimentation, we validate the efficacy of our approach, demonstrating its capability to effectively handle various challenges encountered in real-world surveillance scenarios, such as changes in scale, viewpoint, and environmental conditions. Our findings underscore the importance of integrating multiple feature extraction techniques for robust vehicle re-identification, ultimately contributing to advancements in video surveillance systems.
- Histogram (Color)
- HOG (Histogram of Oriented Gradients)
- SIFT (Scale-Invariant Feature Transform)
- ResNet (ImageNet pre-trained)
- OSNet (torchreid re-id multi-domain dataset pre-trained)
- Composite (SIFT + OSNet)
- License plate recognition
pip install -r requirements.txt
(cs5330) ➜ vehicle-reidentification git:(main) ✗ python main.py -h
usage: main.py [-h] -i INPUT -r REFERENCE_IMG [-f {hog,histogram,sift,resnet,osnet,composite,anpr}] [-anpr ANPR]
options:
-h, --help show this help message and exit
-i INPUT, --input INPUT
path to input video
-r REFERENCE_IMG, --reference_img REFERENCE_IMG
path to reference image
-f {hog,histogram,sift,resnet,osnet,composite,anpr}, --feature {hog,histogram,sift,resnet,osnet,composite,anpr}
Feature type to use for comparison
-anpr ANPR, --anpr ANPR
number plate to match
The percentage value represents the number of frames in which the vehicle was correctly identified out of the total number of frames the vehicle was present in the video. The percentage value is followed by the number of false positives (FP) encountered during the identification process.
(video name in brackets of Vehicle Type, check reference_vehicles/ and videos/ folders.)
| Vehicle Type | Histogram | HOG | SIFT | ResNet | OSNet | Composite |
|---|---|---|---|---|---|---|
| red car (front) | 70% | 70% | 70% | 80% | 80% | 80% |
| white car (front) | 20% | 20% | 70% | 80% (1 FP) | 80% | 85% |
| black car (back) | 80% | 80% | 70% | 95% | 95% | 95% |
| blue car (back) | 80% | 70% | 70% | 95% | 95% | 95% |
| black car (front2) | 100% (2 FP) | 95% | 80% (1 FP) | 100% | 100% | 95% |
| white car (front2) | 70% (2 FP) | 70% (1 FP) | 70% | 90% (2 FP) | 95% | 100% |
| white car (side) | 20% | 50% | 80% (1 FP) | 90% (1 FP) | 90% | 100% (1 FP) |
| grey car (top) | 30% | 50% | 95% | 100% | 100% | 100% |
| black car (top) | 100% | 100% | 100% | 100% | 100% | 100% |
Tested on Macbook Pro M3 8GiB
| Feature | Time (s)/vehicle |
|---|---|
| histogram | 0.002 |
| HOG | 0.026 |
| SIFT | 0.028 |
| ResNet | 0.613 |
| OSNet | 0.240 |
| composite | 0.274 |
| Contributor 1 | Contributor 2 |
|---|---|
| Harshit Kumar | Khushi Neema |