Inter national J our nal of Electrical and Computer Engineering (IJECE) V ol. 16, No. 3, June 2026, pp. 1493 ∼ 1507 ISSN: 2088-8708, DOI: 10.11591/ijece.v16i3.pp1493-1507 ❒ 1493 Hard war e-awar e comparati v e study of lightweight con v olutional neural netw orks f or Raspberry Pi-based autonomous dri ving Hyung In Kim 1 , Y oungmin P ark 2 1 School of Electrical and Electronic Engineering, Y onsei Uni v ersity , Seoul, South K orea 2 Di vision of Con v er gence, Sookmyung W omen’ s Uni v ersity , Seoul, South K orea Article Inf o Article history: Recei v ed Jan 12, 2026 Re vised Feb 2, 2026 Accepted Mar 16, 2026 K eyw ords: Autonomous dri ving Edge AI Embedded deep learning Lightweight con v olutional neural netw ork Raspberry Pi ABSTRA CT Deplo ying deep learning models for autonomous dri ving on resource- constrained edge de vices, such as the Raspberry Pi, presents signicant chal- lenges due to strict limitations on inference latenc y and memory capacity . T o address these constraints, this study conducts a comprehensi v e compara- ti v e e v aluation of lightweight con v olutional neural netw orks (CNNs) optimized for dual-output re gression of steering angle and dri ving speed. W e bench- mark a task-specic end-to-end baseline (NVID IA CNN) ag ainst representa- ti v e classication-oriented architectures—including MobileNet, Shuf eNet, Ef- cientNet, GhostNet, and SqueezeNet—all reformulated for this re gression task. Experiments were conducted on a ph ysical Raspberry Pi-based autonomous RC car platform t o assess prediction accurac y , inference s peed, and real-w orld closed-loop dri ving stability using quantitati v e metrics such as the normalized jerk ratio. Experimental results demonstrate a clear trade-of f: while Ghost- NetV1 0.5 × achie v ed the highest re gression accurac y with a T otal R 2 score of 95.8% and MobileNetV1 recorded a competiti v e MAE of 1.95, the y f ailed to pro vide stable control due to se v ere high-frequenc y steering jitter . Con v ersely , the NVIDIA CNN pro v ed to be the most practical solution for general edge deplo yment, achie ving the lo west inference la tenc y of 61.1 ms (16.4 FPS) and a minimal memory footprint of 2.78 MB, ensuring stable autonomous na vig a- tion (1.50 × jerk ratio). Furthermore, Shuf eNetV2 0.5 × emer ged as the supe- rior archite cture for trajectory precision, recording the lo west weighted MAE of 1.60. These ndings underscore that theoretical accurac y does not guarantee real-w orld dri v ability on em bedded systems, pro viding practical guidelines for hardw are-a w are model selection in edge-based autonomous dri ving. This is an open access article under the CC BY -SA license . Corresponding A uthor: Y oung-Min P ark Di vision of Con v er gence, Sookmyung W omen’ s Uni v ersity 100 Cheongpa-ro 47-gil, Y ongsan-gu, Seoul-04310, South K orea Email: ympillo w@sookmyung.ac.kr 1. INTR ODUCTION Autonomous dri ving has e v olv ed signicantly with the adv ent of deep learning, particularly through end-to-end learning paradigms that map ra w sensor inputs directly to control commands [1], [2]. While high- performance computing platforms ha v e successfully demonstrated autonomous capabilities, deplo ying these systems on resource-constrained edge de vices remains a critical challenge [3]. Embedded platforms, such as J ournal homepage: http://ijece .iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
1494 ❒ ISSN: 2088-8708 the Ras pberry Pi, are widely used for small-scale mobility and deli v ery robots due to their cost-ef fecti v eness and accessibility . Ho we v er , their limited computational throughput and memory capacity impose stringent constraints on e x ecuting con v olutional neural netw orks (CNNs) in real-time. The latenc y induced by computa- tionally intensi v e models directly de grades control stability in closed-loop dri ving scenarios, making hardw are- a w are model selection paramount [4]. T o mitig ate these computational constraints , v arious lightweight CNN architectures—such as Mo- bileNet [5], Shuf eNet [6], and GhostNet [7]—ha v e been proposed. Although these models demonstrate su- perior ef cienc y in image classication tasks, their applicability to re gression-based autonomous dri ving on edge hardw are has not been systematically v alidated. Prior studies on embedded autonomous dri ving ha v e pri- marily relied on simula tion-based v alidation [8] or of ine accurac y metrics [9]. These approaches often f ail to fully capture the critical trade-of fs between inference latenc y , memory usage, and real-w orld dri ving stability . Consequently , there is a lack of comprehensi v e benchmarks that quantitati v ely e v aluate whether classication- oriented lightweight architectures can ef fecti v ely replace task-specic models ( e .g . , NVIDIA CNN) on se v erely resource-constrained platforms. In t his study , we bridge this research g ap by conducting a hardw are-a w are comparati v e e v aluation of lightweight CNN architectures for real-time end-to-end autonomous dri ving. The primary contrib ution of this w ork is the establishment of a v eried, accessibl e Edge-AI testbed using a Raspberry Pi-based platform. As highlighted in recent surv e ys, small-scale v ehicle platforms serv e as a critical bridge for v alidating autonomous systems prior to full-scale deplo yment [10]. Th i s of fers a cost-ef fecti v e alternati v e to full-scale autonomous v ehicles, enabling researchers to v alidate deep learning-based steering and speed prediction algorithms in a real-w orld ph ysical en vironment. Furthermore, we present a unied re gres sion adaptation strate gy to repur - pose di v erse classication-oriented lightweight CNNs into dual-output re gression models, f acilitating a f air and consistent comparison of architectural ef cienc y in edge computing scenarios [3]. Finally , by system- atically analyzing the tr ade-of fs between accurac y , latenc y , and memory usage, we pro vide a comprehensi v e benchmark and research en vironment. This aligns with the emer ging need for ener gy-ef cient Edge AI frame- w orks in autonomous dri ving [11], establis hing a clear reference point for future researchers to implement and quantitati v ely benchmark their o wn algorithmic ideas ag ainst state-of-the-art baselines on resource-constrained hardw are. The remainder of this paper is or g anized as follo ws. Section 2 re vie ws related w ork on end-to-end autonomous dri ving and lightweight CNN architectures. Section 3 describes the system o v ervie w and e xper - imental platform. Section 4 presents the unied model adaptation and implementation strate gy . Section 5 details the e xperimental setup, while section 6 discusses the comparati v e e v aluation results. Finally , section 7 concludes the paper and outlines directions for future w ork. 2. RELA TED W ORKS End-to-end autonomous dri ving using CNNs has been widely studied since NVIDIA demonstrated that s teering angles can be directly re gressed from monocular camera images without e xplicit modular pipelines [1]. This paradigm shift signicantly simplied autonomous dri ving systems and enabled their deplo yment on small-scale platforms. W ith the increasing demand for real-time autonomy on embedded and edge de vices, se v eral studies ha v e e xplored autonomous dri ving on resource-constrained hardw are such as the Raspberry Pi. Projects such as DeepPicar [12] and its v ariants demonstrated that lo w-cost platforms can support CNN-based dri ving, albeit with strict limitations on model comple xity and inference latenc y . More recent studi es [13] further in v estig ated Raspberry Pi-based autonomous RC cars for educational and e xperimental purposes, conrming the feasibility of real-time control under constrained computational b udgets. T o address these constraints, lightweight CNN architectures originally de v eloped for mobile and em- bedded vision ha v e been proposed. MobileNetV1 [5] introduced depth-wise separable con v olution to signif- icantly reduce computational cost and parameter count, while MobileNetV3 [14] further optimized inference latenc y through neural architecture search and hardw are-a w are acti v at ion functions. Shuf eNetV2 [6] focused on minimizing practical memory access cost rather than theoretical FLOPs, achie ving high inference speed on ARM-based CPUs. Ef cientNet [15] and Ef cientNetV2 [16] emplo yed compound scaling and fused con v olu- tion strate gies to impro v e parameter ef cienc y and training speed, respecti v ely . Additional archite ctures such as ResNet [17], GhostNet [7], and SqueezeNet [18] e xplored alternati v e strate gies including residual learn- ing, feature redundanc y reduction, and aggressi v e parameter compression. Although these lightweight models Int J Elec & Comp Eng, V ol. 16, No. 3, June 2026: 1493-1507 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 1495 ha v e demonstrated strong performance in image classication tasks, their applicability to re gression-based au- tonomous dri ving—particularly under real hardw are constraints—remains insuf ciently e xplored. Most prior studies either e v aluated classication accura c y alone or relied on simulation-based benchmarks [8], [10], [19], without comprehensi v e real-w orld dri ving v alidation on edge de vices. In contrast to e xisting w ork, this study conducts a hardw are-a w are comparati v e analysis of multi ple lightweight CNN architectures adapted for dual-output re gression of steering angle and dri ving speed. All models are e v aluated under i dentical e xperimental conditions using a Raspberry Pi-based autonomous RC car , enabling a f air comparison across prediction accurac y , inference latenc y , memory usage, and real-w orld dri v- ing stability . By benchmarking classication-oriented lightweight architectures ag ainst a task-specic base- line (NVIDIA CNN), this w ork aims to identify the most suitable deep learning architecture for real-time autonomous dri ving on se v erely resource-constrained edge platforms. 3. SYSTEM O VER VIEW This section describes the autonomous RC car platform used for real-w orld e v aluation of li gh t weight CNN archit ectures. The system w as designed to pro vi de a reproducible and hardw are-constrai ned en vironment suitable for benchmarking real-time end-to-end autonomous dri ving models. 3.1. Hard war e platf orm The e xperimental platform is a small -scale autonomous RC car equipped with a Raspberry Pi 4 Model B (Broadcom BCM 2711 , 4 GB RAM) as the primary processing unit. Operating at 1.5 GHz, this board serv es as a representati v e resource-constrained edge de vice [20]. A monocular front-f aci ng camera is mounted on the v ehicle to capture road images during operation. Lateral control (steering) is actuated via a s erv o motor , while longitudinal control (propulsion) is pro vided by a DC motor . All actuation and sensor interf acing are handled through the Raspberry Pi’ s general-purpose input/output (GPIO) interf ace. T o ensure stable real-time operation, the system is po wered by an e xternal battery supply capable of sustaining continuous autonomous dri ving without v oltage uctuation. Figures 1(a) and 1(b) present the side and top vie ws of the assembled v ehicle, illustrating the ph ysi cal inte gration of the primary processing unit. The o v erall hardw are connecti vity , detailed in Figure 1(c), delineates the data o w from sensors to the Raspberry Pi and subsequently to the motor dri v ers. This hardw are congu- ration w as intentionally selected to reect the computational and memory limitations commonly encountered in embedded autonomous systems. (a) (b) (c) Figure 1. Ov ervie w of the Raspberry Pi-based autonomous RC car platform used for real-w orld e v aluation, (a) side vie w , (b) top vie w , and (c) hardw are conguration Har dwar e-awar e compar ative study of lightweight con volutional neur al networks for ... (Hyung In Kim) Evaluation Warning : The document was created with Spire.PDF for Python.
1496 ❒ ISSN: 2088-8708 3.2. Softwar e ar chitectur e The s oftw are system w as implemented in Python using open-source libraries. Image acquisition and preprocessing were performed using OpenC V [21], while all deep learning models were implemented and e x- ecuted using the PyT orch frame w ork [22]. The operational w orko w consists of three sequential stages: image data collection, model training and parameter optimization, and real-time autonomous dri ving e x ecution. Dur - ing data collection, camera images were captured and labeled with corresponding steering angles and dri ving speeds obtained through manual operation. F or model training, the collected dataset w as processed of ine using GPU acceleration, and the optimized model parameters were subsequently deplo yed to the Raspberry Pi. During autonomous dri ving, the system operates in a closed-loop manner . At each control c ycle, a cam- era frame is captured, processed by the deplo yed CNN model to predict steering and speed commands, and immediately applied to the v ehicle actuators. This pipeline enables real-time e v aluation of inference latenc y , control stability , and dri ving feasibility under practical hardw are constraints. Figure 2(a) illustrates the o v erall diagram of the softw are w orko w . As sho wn in Figure 2(b), images are captured using a wide-angle camera while the v ehicle is manually controlled with a mouse, and the collected images under go a preprocessing stage. Afterw ard, as illustrated in Figure 2(c), the images are classied, and deep learning is performed either directly on the Raspberry Pi or by compressing the entire dataset and uploading the compressed le to a Google Colab en vironment to generate the deep learning model. Figure 2(d) demonstrates the autonomous dri ving process using the trained deep learning model. (a) (b) (c) (d) Figure 2. Softw are conguration of proposed autonomous RC car , (a) a block dagram of softw are, (b) collecting images for preprocessing, (c) image classication and generating model, and (d) autonomous dri ving with deep learning model 4. MODEL ARCHITECTURE AND IMPLEMENT A TION 4.1. Unied r egr ession adaptation strategy All e v aluated architectures were originally designed for image classication tasks. T o repurpose them for end-to-end autonomous dri ving, we implemented a unied re gression head adaptation scheme, as illus- trated in Figure 3. Each model w as systematically adapted to perform dual-output re gression, simultaneously predicting steering angle and dri ving speed from a single visual input. Int J Elec & Comp Eng, V ol. 16, No. 3, June 2026: 1493-1507 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 1497 Figure 3. Unied re gression head adaptation scheme applied to lightweight CNN backbones First, the original classication heads were remo v ed and replaced with fully connected layers, pro- ducing tw o continuous scalar outputs corresponding to steering and v elocity commands. S econd, all models were congured t o process camera images in a unied input format of 66 × 200 pix els i n the YUV color space, which is consistent with the NVIDIA CNN baseline [1] and optimized for embedded inference. This resolution pro vides suf cient visual information for road feature e xtraction while minimizing computational o v erhead. Third, acti v ation functions, normalization layers [23], and re gularization strate gies were preserv ed as dened in each original architecture, e xcept where minor modications were required to stabilize re gression perfor - mance. Dropout [24] w as selecti v ely applied to fully connected layers to mitig ate o v ertting across all models. This unied adaptation strate gy ensures that performance dif ferences arise from architectural characteristics rather than task-specic implementation bias. 4.2. Baseline ar chitectur e: NVIDIA CNN The NVIDIA CNN serv es as the task-specic baseline for comparison. Unlik e classication-ori ented architectures, this model w as originally designed for end-to-end autonomous dri ving and directly re gresses steering angles from monocular camera input [1]. The architecture, sho wn in Figure 4, consists of v e con v olutional layers follo wed by fully connected layers, emplo ying strided con v olutions for progressi v e spatial do wnsampling. Exponential linear unit (ELU) acti v ation functi o ns , dened as sho wn in (1), are utilized throughout the netw ork to enhance training stability in re gression tasks [25]. Due to its shallo w depth and absence of comple x m odu l es, the NVIDIA CNN e x- hibits minimal memory usage and lo w inference latenc y , making it well-suited for deplo yment on embedded platforms. Figure 4. NVIDIA CNN architecture Har dwar e-awar e compar ative study of lightweight con volutional neur al networks for ... (Hyung In Kim) Evaluation Warning : The document was created with Spire.PDF for Python.
1498 ❒ ISSN: 2088-8708 E LU ( x ) = ( e x − 1 , x ≤ 0 x , x > 0 (1) 4.3. Lightweight CNN ar chitectur es f or edge deployment T o e v aluate the ef fecti v eness of classication-oriented lightweight models in autonomous dri ving re- gression tasks, se v eral representati v e architectures were selected based on their distinct ef cienc y strate gies. This selection aligns with recent ndings that optimized CNNs remain the most viable option for real-time edge inference [26]. MobileNet architectures [5] emplo y depth-wise separable con v olutions to signicantly reduce computational cost. MobileNetV3 [14] further incorporates hardw are-a w are neural architecture search and lightweight acti v ation functions to optimize CPU inference latenc y . Shuf eNetV2 [6] emphasizes practical e x ecution speed on embedded hardw are by m inimizing memory access cost through channel shuf ing mecha- nisms. Ef cientNet [15] and Ef cientNetV2 [16] focus on parameter ef cienc y through compound scaling and fused con v olutional blocks, respecti v ely , while ResNet [17] introduces residual learning to stabilize gradient propag ation in deeper netw orks. GhostNet [7] and SqueezeNet [18] adopt alternati v e lightweight strate gies, including feature redundanc y reduction and aggressi v e parameter compression. Instead of using full-scale architectures intended for serv er -grade GPUs, we selected their compact v ariants (e.g., MobileNetV3-Small [14], Shuf eNetV2 0.5 × [6]) to ensure compatibility with the limited re- sources of the Raspberry Pi. As detailed in section 4.1, the classication heads of all selected backbones were replaced with the unied re gression head sho wn in Figure 3. This ensures that performance v ariations are attrib utable solely to the backbone architecture. 5. EXPERIMENT AL SETUP 5.1. Dataset collection T o e v aluate the performance of lightweight CNN architectures under real-w orld conditions, an image dataset w as collected using the Raspberry Pi-based autonomous RC car on the test track sho wn in Figure 5. The custom test track w as constructed with o v erall dimensions of 3.1 × 1.3 m. This testbed w as iterati v ely designed o v er tw o years to encompass a comprehensi v e set of dri ving primiti v es required for rob ust autonomous na vig a- tion. In contrast to simplistic loop tracks frequently utilized in prior studies [10], our nal track conguration inte grates high-comple xity scenarios, including multiple straight and curv ed se gments with di v erse curv ature proles, comple x intersections (X-crossing) and T -junctions, lane mer ging scenarios (tw o lanes con v er ging into one), right-turn maneuv ers at fork ed paths, and simulated pedestrian crossw alks. This ’all-in-one’ design rigorously tests the model’ s ability to handle di v erse control tasks within a single lap, thereby pro viding a suf- cient le v el of en vironmental comple xity to v al idate the architectural feasibility of the proposed models. Data acquisition w as performed by manually dri ving the v ehicle along the track for v e complete laps, sequentially tra v ersing designated points from A to Z. By completing v e laps on this track, suf cient training data for deep learning model construction w as collected, consisting of track images along with corresponding steering angles and dri ving speeds. Figure 5. Autonomous RC car test dri ving track Int J Elec & Comp Eng, V ol. 16, No. 3, June 2026: 1493-1507 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 1499 A total of 1,647 images were collected from v e complete laps, each paired with corresponding stee r - ing angle and dri ving speed v alues. Steering angles were recorded as inte ger v alues in the range of -99 to +99 with a resolution of 1, while dri ving speeds were recorded in increments of 5 within the range of 40 to 90. All captured images were resized and con v erted to a 66 × 200 YUV format, which w as consistently used across all e v aluated models. The dataset w as di vided into 1,314 training samples and 326 v alidation samples, representing an approximate 80:20 split. Additionally , a distinct subset of 7 samples w as reserv ed specically for qualitati v e visual v erication of model predictions. This identical data split w as applied to all models to ensure f air and consistent e v aluati o n. The detailed specications of the collected dataset are summarized in T able 1. 5.2. T raining conguration T o bypass the computational constraints of the edge de vice, all model training w as performed of f- board using an NVIDIA T4 GPU in the Google Colab en vironment. Optimization w as conducted using the Adam optimizer [27] with an initial learning rate of 1 × 10 − 3 , training for a maximum of 300 epochs. Batch sizes were congured at 16 for training and 64 for v alidation. T o mitig ate o v ertting and conserv e compu- tational resources, an early stopping mechanism w as applied, terminating training when v alidation loss f ailed to impro v e o v er a patience of 30 epochs. The model parameters corresponding to the minimum v alidation loss were retained for subsequent deplo yment. Gi v en that lateral control (steering) stability is paramount for autonomous na vig ation, a weighted mean squared error loss w as adopted. The loss contrib ution w as e xplic- itly weighted at 0.9 for steering angle and 0.1 for dri ving speed. This conguration reects practical dri ving requirements, where minor steering errors may result i n lane de viation, whereas moderate speed errors are generally tolerable. The detailed training h yperparameters are summarized in T able 2. T able 1. Specications of the collected dataset P arameter Specication T otal samples 1,647 images Input resolution 66 × 200 pix els (YUV) Steering label [-99, +99], Inte ger , Resolution: 1 Speed label [40, 90], Inte ger , Step: 5 T raining set 1,314 samples (79.8%) V alidation set 326 samples (19.8%) V isualization set 7 samples (0.4%) T able 2. T raining h yperparameters Hyperparameter Specication Hardw are NVIDIA T4 GP U (Google Colab) Optimizer Adam ( β 1 = 0.9, β 2 = 0.999 ) Learning rate 1 × 10 − 3 Batch size 16 (T rain) / 64 (V alidation) Max Epochs 300 Early stopping P atience: 30 Epochs Loss weights Steering: 0.9, Speed: 0.1 5.3. Deployment and e v aluation metrics The deplo yment w orko w follo wed a cloud-to-edge paradigm. Data collection and preproc essing were performed on the edge de vice, follo wed by model training on a high-performance cloud GPU en vironment (Google Colab) to ensure con v er gence stability . Ho we v er , for the inference phase on the autonomous v ehicle, we intentionally restricted the e x ecution to the Raspberry Pi CPU without utilizing hardw are accelerators ( e .g . , Edge TPU or GPU). This constraint w as imposed to: i) Quantify the intrinsic architectural ef cienc y of the lightweight models without hardw are-specic optimizations masking their computational o v erhead; and ii) V erify the feasibility of autonomous dri ving on ubiquitous, lo w-cost commodity hardw are, thereby establishing a baseline for entry-le v el edge deplo yment. T o assess real-time feasibility , tw o temporal performance metrics were measured during clos ed-loop dri ving: i) inference time, dened as the time required to process a single image and generate steering and speed predictions; and ii) frame time, dened as t he tot al duration of a control c ycle, including image capture, inference, and preparation for the ne xt frame. T o mitig ate e xperimental v ariance, each model w as e v aluated o v er three consecuti v e laps on the tes t track, and a v erage v alues were reported. In addition to these real-time latenc y measurements, memory usage, model size, parameter count, and prediction accurac y on the v alidation set were recorded to pro vide a com- prehensi v e assessment of deplo yment feasibility on resource-constrained hardw are. Furthermore, to pro vide a single holistic metric for model performance, we introduced the weighted total R 2 score. This score represents a weighted re gression coef cient combining steering angle and speed predictions, utilizing the same 0.9/0.1 weighting scheme as the training loss function described in section 5.2 . This weighted aggre g ation reects the prioritization of steering stability o v er speed control in the autonomous dri ving task. T o e v aluate the quality of autonomous na vig ation, we utilized interv ention rate and smoothness (jerk) as practical robotics metrics. T o maintain e xperimental consistenc y and objecti vity , a strict zero-human- Har dwar e-awar e compar ative study of lightweight con volutional neur al networks for ... (Hyung In Kim) Evaluation Warning : The document was created with Spire.PDF for Python.
1500 ❒ ISSN: 2088-8708 interv ention protocol w as enforced during all test runs. Consequently , the interv enti on rate w as recorded as 0 for all e v aluated models, as an y track de viation resulted in the immediate termination of the run (f ailure) rather than manual correction. Re g arding stability analysis, we formul ated a composite metric termed the weighted normalized jerk ratio. First, ra w Jerk v alues—dened as the rate of change in control commands—were mea- sured independently for both steering angle and dri vi n g speed across all models. Crucially , the corresponding jerk v alues generated by the human dri v er were also recorded to serv e as the baseline for smoothness. T o deri v e a relati v e smoothness score, the r a w jerk v alues of each model were normalized ag ainst the human’ s baseline, ef fecti v ely calculating a ratio of model-to-human beha vior . Subsequently , these normalized steering and speed ratios were aggre g ated using the specic 0.9/0.1 weighting scheme established in our training conguration (section 5.2). This metric quanties ho w closely each model replicates the stable dri ving characteristics of a human operator . Similarly , to assess the aggre g ate predicti v e error with a focus on lateral control, we utilized the weighted mean absolute error (weighted MAE), using the same weighting scheme (section 5.2). 6. RESUL T AND DISCUSSION 6.1. T raining r esults of lightweight models All e v aluated model s were trained using an identical dataset and opti mization protocol to ensure f air comparison. As illustrated in Figure 6, most architectures e xhibited rapid loss reduction during the initial training phase, follo wed by gradual con v er gence. The minimal g ap observ ed between training and v alidation loss curv es indicates that se v ere o v er tting did not occur across the e v aluated models. (a) (b) (c) (d) (e) (f) (g) (h) (i) Figure 6. T raining and v alidation loss curv es of e v aluated models, (a) NVIDIA CNN, (b) MobileNetV1, (c) MobileNetV3-Small, (d) Shuf eNetV2 0.5 × , (e) Ef cientNet-B0, (f) Ef cientNetV2-Small, (g) ResNet-20 (Option B), (h) GhostNetV1 0.5 × , and (i) SqueezeNet Int J Elec & Comp Eng, V ol. 16, No. 3, June 2026: 1493-1507 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 1501 Among the candidates, the NVIDIA CNN in Figure 6(a) demonstrated the most stable con v er gence beha vior , achie ving the lo west minimum v alidation loss despite its relati v ely simple architecture. This result suggests that its structural design is well aligned with re gression-based autonomous dri ving tasks. Similarly , MobileNetV1 in Figure 6(b) sho wed a rapid initial decline in loss with steady con v er gence, maintaining a narro w g ap between training and v alidation sets. In contrast, MobileNetV3-Small in Figure 6(c) and Shuf- eNetV2 0.5 × in Figure 6(d) sho w noticeable oscillations in their v alidation loss curv es during later training stages. These uctuations s uggest sensiti vity to specic data samples or suboptimal optimization dynamics, which ma y introduce minor steering instability during real-w orld dri ving. Ef cientNet-B0 in Figure 6(e) and Ef cientNetV2-Small in Figure 6(f) both con v er ged rapidly and required fe wer epochs; ho we v er , their nal v alidation loss remained slightly higher than that of the baseline model. Similarly , ResNet-20 (Option B) in Figure 6(g) and GhostNetV1 0 . 5 × in Figure 6(h) e xhibited early con v er gence b ut were prone to minor spik es in v alidation loss, indicating some v olatility in learning dynamics. Finally , Squeezenet in Figure 6(i) e xhibited smooth con v er gence with minimal uctuation, indicating consistent parameter updates throughout training. Ov erall, the training results indicate that architectural simplicity and task-specic design play a critical role in stabilizing re gression learning under limited data conditions. 6.2. Pr ediction perf ormance on test dataset T able 3 compares the predicted steering angles and dri ving speeds produced by each trained m o de l ag ainst the ground truth v alues from the estimation dataset. All models demonstrated comparable predicti v e accurac y , successfully reproducing human-lik e dri ving beha viors such as decel erating during curv es and accel- erating on straight se gments . F or samples in v olving lar ge steering angles, predicted v alues closely matched the ground truth across all architectures. In contrast, minor discrepancies were observ ed in near -straight dri ving se gments, where se v eral models slightly underestimated speed and e xhibited a small leftw ard steering bias. This beha vior suggests conserv ati v e control strate gies learned from the dataset. A notable observ ation w as that all models predicted lar ger steering angles and lo wer speeds for a spe- cic sample corresponding to a gentle curv e. This consistent beha vior indicates that the models interpreted subtle visual features as indicati v e of increased curv ature, ef fecti v ely embedding a safety-oriented dri ving pol- ic y . These results conrm that lightweight CNN architectures can capture essential dri ving dynamics when adapted to re gression tasks, e v en under limited training data. T able 3. Comparison of ground truth and predicted angles and speeds on selected test samples Models Images Gr ound T ruth Angle Speed Angle Speed Angle Speed Angle Speed Angle Speed Angle Speed Angle Speed -48 50 -33 55 -8 85 -1 90 +8 75 +38 50 +48 50 NVIDIA CNN -50 52 -29 56 -11 86 -8 84 +20 65 +40 49 +42 48 MobileNetV1 -51 54 -27 52 -10 81 -7 78 +18 68 +43 51 +44 50 MobileNetV3-Small -50 49 -33 53 -11 85 -13 77 +19 67 +43 52 +42 47 Shuf eNetV2 0.5 × -54 53 -26 52 -2 82 -11 81 +15 67 +42 51 +43 49 Ef cientNet-B0 -48 53 -25 54 -4 83 -5 77 +12 73 +38 49 +41 47 Ef cientNetV2-Small -50 52 -28 50 -9 78 -5 82 +19 70 +41 53 +39 50 ResNet-20 (Option B) -51 53 -31 52 -10 78 -5 88 +19 65 +46 53 +36 50 GhostNetV1 0.5 × -51 57 -27 60 -8 86 -10 85 +11 75 +42 51 +42 51 SqueezeNet -49 54 -25 56 -11 80 -2 79 +21 69 +40 51 +40 48 6.3. Real-time perf ormance on Raspberry Pi The quantitati v e e v aluation results, encompassing both on-board performance metrics and of ine pre- dicti v e accurac y , are summarized in T able 4. Notably , with the e xception of ResNet-20, all models achie v ed a 100% autonomous completion rate, successfully na vig ating the test track without an y human interv ention. Although ResNet-20 w as selected as a representati v e shallo w residual netw ork to reect minimal residual ar - chitectures suitable for embedded deplo yment, it recorded a 0% completion rate, f ailing to nish a single lap. This f ailure is primarily attrib uted to its cr itical latenc y and memory inef cienc y . As sho wn i n T able 2, it recorded the highest frame time of 269.5 ms (3.7 FPS), causing se v ere phase lag that pre v ented real-time steer - ing corrections. Furthermore, despite its lo w parameter count, the hea vy memory usage (49.33 MB) induced by residual skip connections created a signicant bandwidth bottleneck on the CPU. This computational b urden, combined with the lo west predicti v e accurac y , resulted in the v ehicle’ s inability to maintain the track trajectory . Har dwar e-awar e compar ative study of lightweight con volutional neur al networks for ... (Hyung In Kim) Evaluation Warning : The document was created with Spire.PDF for Python.
1502 ❒ ISSN: 2088-8708 T able 4. Quantitati v e e v aluation of inference latenc y , memory footprint, accurac y , and dri ving stability on Raspberry Pi Models Frame T ime (ms) ↓ Inference T ime (ms) ↓ Latenc y (FPS) ↑ P arams (M) T otal R 2 ↑ Memory Usage (MB) V alidation Loss ↓ W eighted MAE ↓ W eighted Normalized Jerk Ratio ↓ NVIDIA CNN 61.1 35.9 16.4 0.25 94.3% 2.78 0.19% 2.04 1.50 × MobileNetV1 122.2 95.7 8.2 0.81 92.3% 35.34 0.24% 1.95 2.06 × MobileNetV3-Small 92.9 67.1 10.8 1.51 91.7% 16.77 0.22% 2.21 1.58 × Shuf eNetV2 0.5 × 87.0 60.8 11.5 0.34 93.7% 9.12 0.20% 1.60 1.68 × Ef cientNet-B0 195.5 170.2 5.1 4.01 94.7% 66.79 0.24% 2.30 1.75 × Ef cientNetV2-Small 195.0 187.5 5.1 6.05 93.7% 75.54 0.21% 2.44 1.37 × ResNet-20 (Option B) 269.5 243.3 3.7 0.27 92.9% 49.33 0.26% N/A N/A GhostNetV1 0.5 × 123.8 98.2 8.1 1.31 95.8% 20.37 0.23% 2.18 1.65 × SqueezeNet 101.1 74.9 9.9 0.74 93.0% 24.36 0.21% 2.38 1.60 × Inference latenc y is a critical f actor for real-time autonomous control. The NVIDIA CNN achie v ed the lo west a v erage frame time of 61.1 ms, corresponding to approximatel y 16.4 FPS, outperforming all other models. Shuf eNetV2 0.5 × rank ed second, achie ving around 11.5 FPS. In contrast, Ef cientNet v ariants and ResNet-20 e xhibited frame times approaching or e xceeding 200 ms, resulting in insuf cient throughput for sta- ble control. Memory ef cienc y further distinguished the baseline model. The NVIDIA CNN required less than 3 MB of estimated total memory during runtime, signicantly lo wer than all lightweight classication-oriented architectures. Ev en Shuf eNetV2 0.5 × , the most memory-ef cient among the lightweight models, required o v er three times more memory than the baseline. Although GhostNetV1 0.5 × achie v ed the highest o v erall pre- dicti v e performance (T otal R 2 score), its inference latenc y and memory consumption were substantially higher than those of the NVI DIA CNN. These results highlight the critical tr ade-of f between predicti v e accurac y and real-time feasibility on embedded hardw are. F or the autonomous dri ving successful models, we rst conducted a visual inspection of dri ving qual- ity using the temporal de viation prole in Figure 7. This graph plots the real-time de viation of the models’ predictions relati v e to the human’ s dri ving record for steering angle and speed, as sho wn in Figures 7(a) and 7(b). Based on these trajectory data, we deri v ed the predicti v e accurac y metric, MAE, as illustrated in Figure 8. Compares the pe rformance of all e v aluated architectures in terms of steering angle MAE and speed MAE pre- sented in Figures 8(a) and 8(b). Quantitati v ely , MobileNetV1 and Shuf eNetV2 0.5 × demonstrated superior performance, achie ving the lo west Speed MAE (2.38) and Steering MAE (1.51), respecti v ely . (a) (b) Figure 7. Real-time de viation proles of (a) steering angle and (b) speed Int J Elec & Comp Eng, V ol. 16, No. 3, June 2026: 1493-1507 Evaluation Warning : The document was created with Spire.PDF for Python.