Ⅰ. Introduction
Autonomous delivery robots are increasingly used for last-mile transport in residential complexes and pedestrian-shared environments (Lee, 2022;Jeon and Hwang, 2020). They operate in indoor corridors and outdoor pedestrian areas while traversing ramps, tactile paving, and uneven surfaces near people and fixed obstacles. These conditions can produce physical contact and large inertial responses during normal operation. Reliable monitoring must therefore detect actual contact without treating routine disturbances as collisions. Deliberate kicks create additional ambiguity because they resemble obstacle collisions but have a different operational meaning. We therefore treat human interference as a collision-like hard-negative class for collision reporting. Traditional collision detectors apply thresholds to acceleration, angular velocity, wheel speed, or commanded-to-measured motion discrepancies (Cho et al., 2014;Huang et al., 2025). These methods are efficient but sensitive to threshold selection. A sensitive threshold responds to surface vibration and abrupt maneuvers, whereas a conservative threshold can miss weaker collisions. Signal magnitude also varies with speed, payload, traction, and surface condition. Adaptive thresholds reduce this variability, but handcrafted rules still lack the temporal context needed to interpret ambiguous impacts. Learning-based methods capture nonlinear and temporal patterns from sensor sequences (Jo and Kwon, 2020;Seo and Jung, 2023). However, many rely on dedicated contact sensors or constrained experimental settings. Dedicated hardware increases system cost and complicates integration on compact delivery robots. Existing methods also seldom combine intended motion, LiDAR-inertial state estimates, impact indicators, and environmental cues within one event-level decision process. A practical system should combine these cues using sensors already available for robot navigation and perception. In this paper, we present a hybrid multi-sensor collision detection framework for autonomous delivery robots operating in indoor and outdoor pedestrian areas. Physical collision is the target event, while normal driving and deliberate human interference are modeled as alternative states.
The proposed framework comprises two components. (A) A motion-residual module computes translational and rotational discrepancies between commands and Direct LiDAR-Inertial Odometry (DLIO) estimates. It also computes a forward-velocity discrepancy cue and evaluates these indicators using adaptive thresholds. (B) A Temporal Convolutional Network (TCN) classifier fuses control commands, DLIO odometry, filtered impact indicators, and a depth-based person-proximity cue. Temporal refinement converts frame-wise predictions into stable event decisions and separates collisions from normal driving and collision-like human interference. The framework requires no dedicated collision sensor.
The main contributions are as follows:
-
We formulate motion-discrepancy indicators for a wheeled delivery robot by combining commanded-to-DLIO residuals with a forward-velocity cue. This formulation provides motion-aware collision evidence across changing operating conditions.
-
We introduce a multi-sensor TCN classifier with temporal refinement and use collision-like human interference as an explicit hard-negative class.
-
We evaluate the framework across indoor, outdoor, and mixed environments using an adaptive rule-based baseline, recurrent models, and input-feature ablations.
Experiments used approximately 26,000 synchronized frames collected at 10 Hz on a Ranger Mini v1.0. Training used 53 labeled events from one session, while validation used 111 events from six independent sessions. The validation set included 55 collisions and 56 human-interference cases. The TCN achieved an F1-score of 0.89, compared with 0.79 for RNN and 0.86 for LSTM. The adaptive rule-based comparison also indicated fewer false collision decisions under the tested conditions. Together, these results support collision monitoring using existing onboard signals
Ⅱ. Related Works
1. Threshold and Sensor-based Collision Detection
Traditional collision detection methods primarily rely on threshold-based criteria applied to sensor measurements such as acceleration, angular velocity, wheel speed, or motion discrepancies. Owing to their low computational cost and ease of implementation, these approaches have been widely adopted in real-time robotic control systems. For instance, Cho et al.(2014) proposed a velocity-error-based collision detection algorithm that monitors deviations between commanded and measured velocities in real time to identify external impacts. More recently, Huang et al.(2025) proposed an adaptive collision detection method for drones that compares expected motion responses with onboard IMU measurements of linear acceleration and angular velocity using dynamically adjusted thresholds.
Beyond direct sensor-threshold approaches, kinematic and perception-based methods have also been explored to improve collision detection and prediction. Kim et al.(2025) developed a time-interval-based collision detection method for four-wheel-independent-steering (4WIS) mobile robots operating in human-shared indoor environments. Their approach integrates kinematic-based robot trajectory calculation, LiDAR-based human detection, and Kalman filter-based prediction to assess potential robot-human collisions.
Despite their simplicity and efficiency, rule- and sensor-based collision detection methods still have several limitations. Sensor noise, uneven surfaces, or abrupt motion changes can lead to false detections, while subtle collisions may be difficult to distinguish from normal motion disturbances. Moreover, their performance can depend strongly on predefined models, thresholds, or operating conditions, limiting their adaptability in complex real-world environments. As a result, recent research has increasingly shifted toward learning-based and context-adaptive frameworks capable of capturing collision dynamics beyond hand-crafted detection rules.
2. Learning-based and Data-Driven Approaches
To address the limitations of static threshold systems, recent studies have applied machine learning and deep learning models to capture nonlinear and contextual patterns in collision-related sensor data (Jo and Kwon, 2020;Seo and Jung, 2023;Park et al., 2024). These approaches aim to learn temporal dependencies and sensor relationships that reflect robot motion, thereby enabling more adaptive recognition of abnormal situations.
-
1. Several torque-based learning frameworks have been introduced to classify collision events or estimate contact-related quantities directly from sensor measurements. Jo and Kwon(2020) compared multiple machine-learning classifiers using joint torque sensor data for collision detection. Similarly, Park et al.(2024) proposed a Transformer-based collision detection model that estimates joint torque from joint-state information and detects collisions based on the difference between estimated and actual torques without requiring force-torque or tactile sensors.
Beyond torque-based collision detection, other data-driven approaches have incorporated navigation and multimodal sensory information for collision and anomaly detection. Seo and Jung(2023) developed an LSTM-based collision prediction and detection framework that jointly analyzes LiDAR-derived relative distance and acceleration time-series data. Cho et al.(2021) proposed an anomaly-detection approach for marine robots that learns environmental flow effects and identifies abnormal speed changes while reducing false alarms caused by unexpected ocean flows. Yang et al.(2025) further introduced an unsupervised anomaly-detection framework based on Mahalanobis SVDD that integrates IMU and audio signals to identify anomalous robot conditions.
Another important distinction in this line of research lies in the sensing requirement. Many prior studies on robot collision or contact detection rely on dedicated sensing hardware, such as torque sensors, force sensors, tactile bumpers, or additional acoustic modules. While these sensors can improve sensitivity, they also increase system complexity and reduce practicality for compact mobile robot platforms. In contrast, our work focuses on detecting collisions and abnormal interference using only the onboard sensors already required for autonomous navigation, including IMU, LiDAR-based state estimation, wheel odometry, and control commands. This makes the proposed framework more suitable for real-world delivery robots, where minimizing additional hardware is an important practical consideration.
3. Motion-Based and Adaptive Collision Detection
Recent research has also explored motion-based approaches that analyze a robot’s kinematic behavior to identify potential collision events or abnormal interactions. Rather than solely relying on thresholding sensor values, these methods use trajectory, velocity, or dynamic-response information to assess potential collision risks and adapt robot motion accordingly (Zardykhan et al., 2019).
Zardykhan et al.(2019) proposed a collision-preventing phase-progress control framework for human-robot collaboration. The method predicts possible robot-human collisions and continuously modulates the robot’s velocity according to the estimated collision risk while preserving the predefined motion path. This approach demonstrates the importance of incorporating motion and interaction context into robotic safety mechanisms.
By incorporating motion cues into collision-risk assessment, these methods provide additional context for potentially unsafe or context-dependent interactions. Motion-adaptive strategies therefore complement threshold-based and learning-based methods by providing additional dynamic context for collision recognition.
Existing studies on robot collision and anomaly detection have progressed from threshold-based sensing to learning-based modeling and motion-aware interpretation. Threshold methods are lightweight and computationally efficient, but they are vulnerable to terrain changes, sensor noise, and subtle impacts. Learning-based approaches improve adaptability by capturing nonlinear sensor patterns, but many rely on specific sensing modalities, dedicated collision-related hardware, or controlled experimental conditions. Motion-based techniques use kinematic discrepancies as useful cues, yet their performance depends on reliable state estimation and is still sensitive to environmental disturbances.
In addition, many prior works primarily focus on collision detection itself, rather than the broader problem of safety monitoring in mobile delivery robots operating in mixed real-world environments. In contrast, this study targets not only physical collisions but also abnormal interference, such as intentional human tampering, using only the sensors already available for robot navigation and motion estimation. To address this gap, we propose a context-aware framework that combines dynamic thresholding with a deep learning-based temporal model using IMU, DLIO (LiDAR-Inertial Odometry), odometry, and control-command information. This design enables practical and deployable recognition of collisions and abnormal interactions without the need for additional collision-specific sensing hardware.
Ⅲ. Methodology
1. Dynamic Threshold-Based Collision Detector
The proposed Dynamic Threshold-Based Collision Detector extends the adaptive collision detection strategy originally developed for aerial robots (Huang et al., 2025) to wheeled delivery robot. To account for the different vehicle dynamics, the original thrust-deviation formulation is replaced with DLIO-based motion estimation and longitudinal slip estimation, enabling robust collision detection under ground operating conditions.
1) DLIO-Based Motion State Estimation
The motion state of the mobility is estimated using the Direct LiDAR-Inertial Odometry (DLIO) framework (Chen et al., 2023), which fuses LiDAR point clouds and IMU measurements to provide accurate six-degree-of-freedom pose and velocity estimation. Through scan-to-map registration and factor graph optimization, DLIO ensures drift-free localization even during extended operation, as illustrated in <Fig. 2>. The estimated linear velocity expressed in m/s, and angular velocity expressed in rad/s are used as observation inputs for the collision detection module.
2) Control Input Acquisition
The control input initially consist of target linear and angular velocities . To ensure high sensitivity to impulsive forces, these velocity commands are further processed into expected longitudinal acceleration and angular acceleration based on the Ackermann steering kinematic constraints. This differential approach allows the system to establish a reference for the mobility’s intended dynamic behavior. By monitoring the residuals between these command-derived accelerations and the actual states estimated by DLIO, the algorithm can effectively isolate high-frequency physical impacts from steady-state cruising. This transformation is pivotal for distinguishing genuine collision-induced shocks from planned maneuvers, such as rapid deceleration or sharp cornering, thereby minimizing false positives in complex traffic environments.
3) Slip Estimation and Compensation Term
To account for wheel-road interaction effects, a slip term is introduced. It measures the discrepancy between commanded and estimated forward velocities:
Here, the slip term represents the absolute difference between the commanded forward velocity and the forward velocity estimated by DLIO, expressed in meters per second.
When this difference exceeds a context-dependent threshold , which represents the allowable velocity discrepancy under the current driving condition, the system infers that significant wheel slippage or traction loss is occurring. While recent collision detection models for aerial vehicles rely on thrust deviation to capture sudden impacts (Huang et al., 2025), ground mobilities are fundamentally governed by friction-dependent kinematics. Therefore, to accurately model wheeled dynamics on physical roads, this study effectively replaces the aerial thrust deviation concept with the proposed longitudinal slip term.
4) Dynamic Threshold-Based Collision Decision
The final decision stage integrates multiple dynamic checks using adaptive thresholds. The algorithm evaluates both translational and rotational consistency by calculating the absolute differences between the commanded and estimated longitudinal accelerations and between the commanded and estimated angular accelerations:
The acceleration-based indicator represents the longitudinal acceleration discrepancy in m/s2, while the yaw-based indicator represents the rotational acceleration discrepancy in rad/s2. These indicators quantify the deviation between the motion expected from the control commands and the actual motion estimated by DLIO.
A collision flag is triggered when either or exceeds its dynamic threshold or when indicates abnormal traction deviation. This multi-criteria design enhances robustness to noise and terrain variation, enabling reliable detection of external collisions even under dynamic driving conditions.
2. Deep Learning-Based Collision Detection and System Integration
The dynamic threshold-based detector introduced in Section 3.1 provides effective rule-based collision identification. However, it inherently lacks adaptability to contextual variations such as changes in driving surface, slope, payload, or environmental vibration. Fixed analytical thresholds cannot fully represent the nonlinear relationships between control inputs, odometry feedback, and sensor responses under diverse real-world conditions. To address these limitations, we designed a deep learning-based collision classification module that learns temporal and nonlinear dependencies directly from multi sensor data sequences. The network receives fused inputs from commanded velocities, odometry estimates, inertial signals, and LiDAR-derived depth information, as summarized in <Table 1>.
<Table 1>
Deep Learning Input data
| Feature Name | Source | Description |
|---|---|---|
| cmd_linear_x | Control input | Commanded linear velocity |
| cmd_angular_z | Commanded angular velocity | |
| odom_linear_x, y, z | DLIO odometry | Actual linear velocity components estimated by DLIO |
| odom_angular_x, y, z | Actual angular velocity components estimated by DLIO | |
| H_acc | IMU /Kalman filter | Kalman-filtered acceleration-based threshold index |
| H_yaw | Kalman-filtered yaw-rate-based threshold index | |
| median_depth | LiDAR / YOLOv11 | Median depth of LiDAR points within YOLOv11 person bounding box |
1) Input Feature Composition
The input features for the proposed deep learning-based collision detection model were designed to represent both the robot’s commanded motion and its observed physical responses. They include control commands, odometry estimates, inertial indicators, and environmental perception cues, allowing the network to model the relationship between intended and actual motion. The control commands (cmd_linear_x, cmd_angular_z) describe the robot’s target linear and angular velocities, while the odometry feedback (odom_linear_x,y,z, odom_angular_x,y,z) from the Direct LiDAR-Inertial Odometry (DLIO) module reflects its real motion state, including deviations such as slip or drift. The dynamic threshold indices (H_acc, H_yaw) quantify acceleration and yaw-rate-based deviations filtered through a Kalman process, providing impact-sensitive motion cues. The environmental feature(median_depth) represents the median LiDAR depth within YOLOv11-detected person regions, indicating obstacle proximity(Fig. 3). These inputs enable the model to learn spatiotemporal dependencies between control, motion, and environmental conditions, forming a unified representation for reliable collision classification.
2) Dataset Configuration
The proposed framework is trained using three representative operational states: normal driving, physical collision, and human tampering. Each sample consists of synchronized IMU, LiDAR, odometry, and control signals sampled at 10 Hz, and is segmented into fixed-length temporal windows before being provided as input to the model. The synchronized multi sensor sequences enable the network to learn temporal motion patterns associated with normal operation, collision events, and external human interference. Detailed descriptions of the data collection procedure, experimental environment, dataset composition, and data split are provided in Section IV.
3) Deep Learning Architecture for Multi sensor Collision Recognition
The proposed collision recognition framework adopts a Temporal Convolutional Network (TCN) architecture to model sequential dependencies in multimodal sensor data. Unlike static threshold methods that rely on instantaneous signals, the TCN captures both short-term and long-term temporal patterns from synchronized control, odometry, and perception inputs, allowing robust identification of contextual transitions such as collisions, abrupt impacts, or external interference. As illustrated in <Fig. 5>, the model receives fused input features consisting of commanded velocities, DLIO-estimated odometry, Kalman-filtered inertial indicators, and LiDAR-based environmental cues. The input sequence is first transposed and passed through a stack of four residual temporal convolutional blocks, each composed of a pair of 1D dilated convolutions (kernel size = 3), ReLU activation, and dropout layers. The dilation rates [1,2,4,8] progressively enlarge the receptive field to cover both local and global temporal dynamics, enabling the model to capture subtle patterns preceding or following impact events. Each residual block preserves temporal resolution through zero-padding and skip connections, stabilizing gradient flow and maintaining alignment between input and output sequences. The encoded features are then mapped to frame-wise class logits through a convolutional classification head, producing a per-frame probability vector across three classes: normal driving, collision, and abnormal interference. A softmax function converts these logits into class probabilities for subsequent post-processing.
This architecture effectively integrates multiple sensing modalities and their temporal correlations without relying on recurrent structures. By using dilated convolutions with residual connections, the network captures both short- and long-range temporal dependencies while maintaining stable gradients and efficient feature propagation. The resulting per-frame predictions serve as the basis for event level refinement described in Section 3.2.4.
4) Post-Processing and Event Level Decision Fusion
After frame-level predictions were produced by the TCN-based classifier, a post-processing stage was applied to convert noisy frame-wise outputs into stable and interpretable event level decisions. First, a temporal hysteresis filter with separate entry and exit thresholds and a patience mechanism was employed to suppress transient oscillations in class probabilities, ensuring that only sustained confidence changes triggered state transitions. Next, the predicted label sequence was refined by merging or removing short and fragmented segments based on their duration, mean probability, and physical consistency with rule-based impact indicators. Gaps between identical events were filled to prevent artificial fragmentation, while rapidly alternating class-1 and class-2 sequences were unified according to their dominant probability trend. Finally, a decision fusion step aligned predicted intervals with post-processing pipeline, the system achieved robust temporal consistency, significantly reduced false alarms, and established a one-to-one correspondence between actual and detected events, thereby enhancing the overall reliability and interpretability of the results. The complete workflow of the proposed hybrid collision detection framework, integrating dynamic threshold, deep learning classification, and post-processing fusion, is summarized in <Fig. 4>.
Ⅳ. Experiments
1. Robot Platform and Sensor Configuration
The primary objective of this experimental setup is to enable reliable collision detection and classification research for autonomous delivery robots operating in both indoor and outdoor environments, particularly in confined areas such as narrow corridors, ramps, and obstacle-rich pedestrian zones. To support data collection and validation in such spaces, we utilize the Ranger Mini v1.0, a compact four-wheeled electric mobile platform designed for agile navigation and Ackermann steering control. Its small footprint and high mobility make it well suited for evaluating robot performance in real-world delivery scenarios that involve frequent direction changes and complex spatial layouts. An overview of the platform and its sensor configuration is shown in <Fig. 5>. The platform is equipped with a multi sensor suite that provides high-frequency motion and environmental data essential for collision monitoring. The configuration is designed with simplicity and robustness in mind, focusing on practical deployability and sensor complementarity for time-series data collection. We employ the Ouster OS1-32 LiDAR, a 32-beam rotating 3D sensor that offers a 360° horizontal field of view at 10 Hz or higher. It generates dense 3D point clouds in the format(x,y,z,intensity), effectively capturing geometric features such as walls, obstacles, and surrounding structures. The LiDAR is mounted at an elevated position on top of the robot to minimize occlusion while maximizing spatial coverage.
An Xsens MTi-430 IMU is used to provide six-axis inertial data, including 3-axis acceleration and 3-axis angular velocity. These signals capture fine-grained dynamic responses of the robot and are particularly important for detecting impact-related variations in acceleration and angular motion. Additionally, an Intel RealSense D435 RGB-D camera is installed on the front of the platform. It captures RGB and depth information to enhance perception in near-field environments, supporting pedestrian detection and short-range distance estimation tasks. All sensor data streams are synchronized and timestamped at consistent intervals, forming a continuous time-series dataset used for model training and evaluation. This integrated sensor configuration enables the collection of multimodal data that supports both dynamic-threshold collision detection and deep-learning-based classification across a wide range of real-world delivery robot environments.
2. Dataset Configuration and Experimental Setups
The dataset used in this study was collected from real-world operation of the autonomous delivery robot in mixed indoor and outdoor environments. Each recorded sequence contained synchronized data from IMU, LiDAR, odometry, and control command topics sampled at 10 Hz, producing approximately 26K frames in total.
The labeled dataset consisted of two abnormal categories, collision and tampering, in addition to normal driving data. As summarized in <Table 3>, the training dataset contained 53 labeled events, while the validation dataset included 111 labeled events, comprising 56 tampering cases and 55 collision cases. The training dataset was collected during a single recording session, whereas the validation dataset was constructed from six independent recording sessions conducted under diverse operating conditions, as summarized in <Table 2>. Consequently, the validation set contains a larger number of events and normal driving frames, enabling a more comprehensive evaluation of the proposed method under varied real-world scenarios. Each event was defined as a discrete detection instance, counted by occurrence rather than by continuous frame length. The normal sequences were included to maintain contextual balance and support temporal modeling. Representative temporal patterns of the synchronized sensor features for the three operational scenarios (normal driving, collision, and tampering) are shown in <Fig. 6>, illustrating the distinct motion characteristics learned by the proposed model.

<Fig. 6>
Adaptive threshold behavior under collision and tampering scenarios
(a) Commanded and actual linear velocities with shaded regions indicating ground-truth events: red for collision and yellow for tampering. In both cases, the acceleration-based impact index occasionally exceeds the adaptive threshold. However, not all exceedances correspond to actual collisions. (b) IMU-derived impact index and its adaptive threshold estimated by the proposed Kalman-based model. The threshold dynamically adjusts according to motion and surface variations, reducing false detections from transient accelerations while retaining sensitivity to true impact events
<Table 2>
Environmental characteristics of the six validation sessions
| Session | Environment | Pavement | Illumination | Spatial Condition | Pedestrian interaction |
|---|---|---|---|---|---|
| 1 | Mixed | Uneven | Daylight | Mixed | Variable (Low–High) |
| 2 | Indoor | Smooth | Daylight | Corridor | Low |
| 3 | Oudoor | Uneven | Night | Open area | Low |
| 4 | Indoor | Smooth | Daylight | Corridor | Low |
| 5 | Mixed | Uneven | Night | Mixed | Variable (Low–High) |
| 6 | Indoor | Smooth | Daylight | Corridor | High |
<Table 3>
Dataset composition by scenario and subset
| Category | Training set | Validation set | Total |
|---|---|---|---|
| Total labeled events | 53 | 111 | 164 |
| Collision events | 28 | 55 | 83 |
| Tampering events | 25 | 56 | 81 |
| Normal frames | ~1.3k | ~20.7k | ~22k |
| Event frames | ~2.7k | ~1.3k | ~4k |
For training, the data were divided into overlapping fixed-length windows of 1.5s and normalized feature-wise to zero mean and unit variance. For comparison, a rule-based baseline was implemented using the adaptive Kalman-filter threshold model described in Section 3.1. This model estimates dynamic thresholds for acceleration and yaw-rate deviations, serving as a reference to evaluate the performance improvement of the proposed deep learning-based approach in terms of false-alarm reduction and event classification accuracy.
The abnormal scenarios in the dataset consist of two categories: collision and tampering. Collision events correspond to cases in which the robot physically contacts surrounding objects that may realistically cause impacts during operation, such as walls or lampposts. Tampering events refer to intentional human interference, particularly kicking the robot. As illustrated in <Fig. 7>, the dataset includes representative samples collected from these real operating environments. In addition, the data were collected in an alternating chronological manner across scenarios, so that the recorded sequences reflect the temporal order of actual deployment rather than isolated event-wise acquisition.
3. Comparative Evaluation with Rule-Based Method
The proposed collision detection framework was compared with the rule-based baseline described in Section 3.2. The conventional model relied on acceleration and yaw-rate thresholds from the adaptive Kalman filter and often misclassified surface vibration or abrupt speed changes as collisions.
<Table 4> presents an event-level comparison of detection performance between the rule-based baseline and the proposed collision detection framework. All evaluation metrics were computed on an event basis, where each collision or tampering occurrence was treated as an independent detection event. True Collision Detection denotes the percentage of correctly detected collision events, False Collision Detection the percentage of non-collision events incorrectly classified as collisions, Missed Collision the percentage of undetected collision events, and Overall Accuracy the overall event level classification accuracy.
<Table 4>
Comparison of collision detection accuracy between rule-based and proposed framework
| Model | True Collision Detection (%) ↑ | False Collision Detection (%)↓ | Missed Collision (%) ↓ | Overall Accuracy (%) ↑ |
|---|---|---|---|---|
| Rule-based | 100 | 62.37 | 0 | 37.62 |
| Ours | 96.72 | 0 | 3.28 | 96.72 |
The rule-based system achieved a True Collision Detectionrate of 100%, but also produced a high False Collision Detectionrate (62.37%) because any exceedance of the adaptive threshold was treated as a collision. Consequently, although all collision events were detected, frequent false alarms during normal driving reduced the overall accuracy to 37.62%. In contrast, the proposed framework achieved 96.72%overall accuracy with only 3.28%missed collisions, demonstrating its ability to distinguish genuine impacts from environmental disturbances and normal driving maneuvers.
As illustrated in <Fig. 6>, the rule-based detector reacts whenever the acceleration index exceeds its threshold, often misinterpreting normal motion as collision events. In contrast, the proposed framework effectively interprets temporal motion context and filters out non-collision cases.
4. Comparison of Deep Learning Temporal Architectures
To evaluate the effectiveness of the proposed Temporal Convolutional Network (TCN) architecture, we compared its performance against recurrent neural models including a standard Recurrent Neural Network (RNN) and a Long Short-Term Memory (LSTM) network. All models were trained on the same dataset using identical input features and hyperparameter configurations to ensure a fair comparison. As summarized in <Table 5>, the proposed TCN achieved the highest F1-score (0.89), outperforming both RNN (0.79) and LSTM (0.86). While LSTM exhibited strong precision (88.0%), its recall was noticeably lower (85.6%), indicating a tendency to miss certain true collision or interference events. In contrast, the TCN maintained a balanced trade-off between precision (87.2%) and recall (91.9%), effectively capturing both clear and subtle impact patterns. These results demonstrate that the TCN’s dilated convolutional structure enables more effective temporal dependency modeling than sequential recurrent architectures.
<Table 5>
Performance comparison among temporal architectures
| Model | Precision (%)↑ | Recall (%)↑ | F1-score↑ | Inference Time (ms/frame) |
|---|---|---|---|---|
| RNN | 69.4 | 91.9 | 0.79 | 4.8 |
| LSTM | 88.0 | 85.6 | 0.86 | 6.3 |
| TCN (Proposed) | 87.2 | 91.9 | 0.89 | 3.2 |
In addition, <Table 6> presents an ablation study designed to quantify the contribution of each component within the proposed framework. The results indicate that detection performance improves progressively as more informative and stabilized inputs are incorporated. Specifically, Control + DLIO achieved 74.36% overall accuracy, Control + DLIO + Depth improved this to 80.77%, and Control + DLIO + Kalman reached 92.31%. The proposed full framework achieved the best performance at 96.72%, indicating that the performance gain results from the combined effects of DLIO-based state estimation, adaptive thresholding, multi-sensor input representation, and TCN-based temporal modeling.
<Table 6>
Ablation Study on Input Modalities for Collision Classification
| Model | True Collision Detection (%)↑ | False Collision Detection (%)↓ | Missed Collision (%) ↓ | Overall Accuracy (%) ↑ |
|---|---|---|---|---|
| Control + DLIO | 74.36 | 1.69 | 25.64 | 74.36 |
| Control + DLIO + Depth | 80.77 | 4.55 | 19.23 | 80.77 |
| Control + DLIO + Kalman (H_threshold) | 92.31 | 0 | 7.69 | 92.31 |
| Full (Proposed) | 96.72 | 0 | 3.28 | 96.72 |
<Table 7> presents the class-wise performance of the proposed model on the validation set. The model achieved F1-scores of 98.71%, 93.9%, and 88.6% for Normal, Collision, and Tampering, respectively. The Tampering class showed slightly lower recall and F1-score than the Collision class, indicating that tampering events were relatively more difficult to detect. Nevertheless, the model maintained high precision across all three classes.
<Table 7>
Class-wise performance of the proposed model on the validation set
| Class | Precision (%)↑ | Recall (%)↑ | F1-score↑ |
|---|---|---|---|
| Normal | 97.77 | 99.67 | 98.71 |
| Collision | 95.8 | 92.0 | 93.9 |
| Tampering | 91.37 | 87.0 | 88.6 |
Ⅴ. Conclusion and Future work
This study presented a hybrid collision detection framework for autonomous delivery mobilities, integrating DLIO-based motion state estimation, adaptive dynamic thresholding, multi-sensor fusion, and a TCN-based contextual classifier.By combining information from LiDAR, IMU, odometry, and control commands, the proposed framework achieved reliable detection of physical impacts and abnormal interactions such as tampering or theft attempts across pedestrian-mixed indoor-outdoor environments. Experimental results demonstrated that the proposed framework significantly reduced false alarms compared with the adaptive rule-based baseline while maintaining high sensitivity to genuine collision events.
















