Robust Parking Space Detection Using Low-Cost Cameras

Overview

This repository contains the implementation of a robust parking space detection system based on global perceptual understanding. The system leverages convolutional neural networks (CNNs) and a transformer-based module called Global Perceptual Feature Extractor (GPFE) to enhance detection accuracy and robustness against lighting variations. The method has been evaluated on CNREXT, PKLot, and ACPDS datasets and compared with results from the original research paper.

Features

• Low-cost vision-based parking space detection
• CNN integration with the GPFE module for enhanced feature extraction
• Evaluation on three benchmark datasets
• Comparison with state-of-the-art models (ShuffleNetV2, mAlexNet)
• Implementation using PyTorch
• Dataset preprocessing, model training, and evaluation scripts included

Datasets

Three benchmark datasets are used in this study:

1. ACPDS: Contains segmented parking space images with labeled occupancy status.
2. PKLot: A large-scale dataset with various weather conditions and camera angles.
3. CNREXT: Features diverse lighting conditions and includes parking lot annotations.

Each dataset has been preprocessed according to the specifications provided in the reference research paper. The dataset segmentation and augmentation procedures are available in the utils/ directory.

Model Architecture

The parking space classification model consists of:

• ShuffleNetV2 / mAlexNet: Lightweight CNN architectures for image classification.
• GPFE Module: A transformer-based feature extractor designed to enhance generalization in varying lighting conditions.
• Data Augmentation: Random cropping and color transformation for better generalization.
• Loss Function & Optimizer: Uses CrossEntropyLoss with SGD optimizer and a learning rate scheduler.

Model Training

1. ShuffleNetV2

   • Implemented using torchvision.models.shufflenet_v2_x1_0
   • Custom modifications for binary classification.

2. mAlexNet

   • Implemented as per the referenced paper specifications.

Training Parameters

• Batch Size: 64 (Due to resource constraints, 32 was used)
• Optimizer: SGD with Momentum (0.9) & Weight Decay (0.0005)
• Learning Rate: 0.001, decayed every 3 epochs by 0.75
• Epochs: 60
• Data Augmentation: Random cropping and color jittering

Results

Table 1: Effect of Classification Attentive Module on mAlexNet+GPFE Accuracy

Acc(%)	ratio	cam	layers	heads	emb	batch
74.16	2	True	12	8	256	32
80.54	16	True	12	8	256	32
86.98	-	False	12	8	256	32
81.61	8	True	8	8	256	32
85.97	-	False	8	8	256	32

---

Table 2-a: Accuracy, Parameters, and FLOPs Comparison (Implementation Results)

Models	FLOPS	Params (M)	ACPDS (%)	PKLot (%)	CNREXT (%)
mAlexNet	28.61	0.04	86.91	97.47	96.16
+GPFE	794.54	10.38	86.98	96.71	96.49
Delta	+765.93	+10.34	+0.07	-0.76	+0.33
ShuffleNet	197.7	1.26	92.75	98.78	98.99
+GPFE	932.09	11.58	89.26	99.09	98.79
Delta	+734.39	+10.32	-3.49	+0.31	-0.20

---

Table 3-a: Cross-Dataset Evaluation (Implementation Results)

Models	Train Dataset	Test Dataset	Accuracy (%)
mAlexNet	PKLot_Train	CNR-EXT	85.25
+GPFE	PKLot_Train	CNR-EXT	82.46
Delta			-2.79
ShuffleNet	PKLot_Train	CNR-EXT	84.39
+GPFE	PKLot_Train	CNR-EXT	81.75
Delta			-2.64

Table 4-a: Cross-Testing on CNREXT Sub-Datasets (Implementation Results)

CNREXT Sub-Dataset	Train Dataset	Test Dataset	ShuffleNet (%)	+GPFE (%)	Delta (%)
Sunny	CNREXT	Sunny	98.55	98.74	-0.19
Rainy	CNREXT	Rainy	97.18	97.53	-0.35
Overcast	CNREXT	Overcast	98.27	98.48	-0.21

---

Table 5-a: Cross-Testing on PKLot Sub-Datasets (Implementation Results)

PKLot Sub-Dataset	Train Dataset	Test Dataset	ShuffleNet (%)	+GPFE (%)	Delta (%)
PUC_test	UFPR04_train	PUC_test	92.99	92.88	-0.11
UFPR04_test	UFPR04_train	UFPR04_test	99.69	99.59	-0.10
UFPR05_test	UFPR04_train	UFPR05_test	90.51	89.13	-1.38
PUC_test	UFPR05_train	PUC_test	94.00	91.14	-2.86
UFPR04_test	UFPR05_train	UFPR04_test	88.81	88.13	-0.68
UFPR05_test	UFPR05_train	UFPR05_test	99.53	99.53	0.00

---

Table 6-a: Impact of GPFE-CA on Model Performance (Implementation Results)

Models	Train Dataset	Test Dataset	Accuracy (%)
mAlexNet	PKLot_Train	CNR-EXT	85.25
mAlexNet + GPFE without CA	PKLot_Train	CNR-EXT	82.46
mAlexNet + GPFE with CA	PKLot_Train	CNR-EXT	81.58
ShuffleNet	PKLot_Train	CNR-EXT	84.39
ShuffleNet + GPFE without CA	PKLot_Train	CNR-EXT	81.75
ShuffleNet + GPFE with CA	PKLot_Train	CNR-EXT	80.38

---

Table 7: Model Size Comparison After Quantization

Models	TensorFlow Model Size (MB)	TensorFlow Lite Model Size (MB)
mAlexNet	0.342	≈0.1
mAlexNet+GPFE	81	≈20
ShuffleNetV2	10	≈3
ShuffleNetV2+GPFE	91	≈40