A vehicle collision detection method based on an improved two-stream convolutional network
By using an improved dual-stream convolutional network, combined with a lightweight network and a hybrid attention module, the problem of balancing real-time performance and accuracy in vehicle collision detection is solved, achieving efficient and reliable vehicle collision detection that meets automotive-grade requirements and improves the active safety performance of intelligent vehicles.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHANGAN UNIV
- Filing Date
- 2026-01-26
- Publication Date
- 2026-05-26
AI Technical Summary
Existing vehicle collision detection technologies struggle to balance real-time performance, accuracy, and deployment costs. Traditional methods suffer from poor environmental adaptability and high computational complexity, while deep learning methods struggle to meet real-time and power consumption requirements on embedded platforms. Furthermore, feature fusion strategies are insufficient and fail to effectively capture the complete evolution of collision risks.
An improved dual-stream convolutional network is adopted, utilizing the lightweight network MobileNetV3 and the depthwise separable convolutional long short-term memory network (DW-ConvLSTM) for feature extraction and temporal attention modeling. Feature fusion is enhanced by a hybrid attention module, and a lightweight and efficient vehicle collision detection system is constructed by combining an adaptive feature fusion module and a staged training strategy.
It achieves high-precision, low-latency, and easy-to-deploy vehicle collision detection, reduces false alarm rates, meets automotive-grade requirements, and provides excellent real-time performance and adaptability, offering a stable and reliable solution for intelligent vehicle active safety systems.
Abstract
Description
Technical Field
[0001] This invention relates to the field of vehicle collision detection technology, specifically to a vehicle collision detection method based on an improved dual-stream convolutional network. Background Technology
[0002] Vehicle collision detection, as a core component of active automotive safety, has always faced the challenge of balancing accuracy, real-time performance, and deployment cost in its technological evolution. Traditional solutions often rely on physical sensors such as radar or simple visual thresholding methods, which, while offering fast response times, are limited by poor environmental adaptability, high costs, and insufficient recognition capabilities in complex scenes. With the development of deep learning, visual methods such as dual-stream convolutional networks have significantly improved detection potential by fusing spatiotemporal information, but their standard architectures face obvious limitations in automotive scenarios: the computationally complex backbone network struggles to meet the real-time and power consumption constraints of embedded platforms; the ability to model long-term motion patterns is insufficient, making it difficult to capture the complete evolution of collision risks; feature fusion strategies are relatively simple, failing to dynamically balance the contribution of different information; and most studies have not fully considered engineering challenges such as image interference, model compression loss, and false alarm rate control in actual driving.
[0003] Therefore, there is an urgent need for a collision detection solution that can achieve high accuracy, low latency, and stable reliability with limited resources, in order to promote the large-scale application of active safety systems. Summary of the Invention
[0004] To solve the above technical problems, the present invention provides the following technical solution: a vehicle collision detection method and system based on an improved dual-stream convolutional network, comprising the following steps: Step S1: Acquire raw video streams containing the target vehicle in real time using the vehicle-mounted camera; Step S2: Preprocess the original video stream to extract the spatial frame sequence and optical flow frame sequence respectively; Step S3: Input the spatial frame sequence into the spatial flow branch of the improved two-stream convolutional network, and input the optical flow frame sequence into the temporal flow branch of the improved two-stream convolutional network; The spatial flow branch uses the lightweight network MobileNetV3 as the feature extraction backbone and embeds a first hybrid attention module in at least one of its bottleneck layers. The temporal branch includes a feature extraction unit and a temporal attention modeling unit connected in sequence; the feature extraction unit uses the lightweight network MobileNetV3 as its backbone and embeds a second hybrid attention module in at least one of its bottleneck layers. The temporal attention modeling unit is a convolutional long short-term memory network module based on depthwise separable convolution, which is used to perform temporal attention weighting and long-term dependency modeling on the feature sequence output by the feature extraction unit. Step S4: Extract high-level spatial features and high-level spatiotemporal motion features through the spatial flow branch and the temporal flow branch, respectively; Step S5: Perform feature fusion on the extracted high-level spatial features and high-level spatiotemporal motion features to generate a fused feature vector; Step S6: Input the fused feature vector into the classifier and output the probability of the target vehicle colliding or the collision detection result.
[0005] Preferably, the time attention modeling unit specifically performs the following operations: Receive a feature map sequence of length T from the feature extraction unit; The feature map sequence is input into the DW-ConvLSTM layer, where the core computation uses depthwise separable convolution instead of standard convolution to reduce the number of parameters. Extract the hidden state of the DW-ConvLSTM layer at the last time step, or perform adaptive pooling on the hidden states of all time steps to obtain the high-level spatiotemporal motion features weighted by temporal attention.
[0006] Preferably, after the DW-ConvLSTM layer, a lightweight spatiotemporal attention submodule is connected to perform secondary enhancement on the hidden state in the channel and spatial dimensions.
[0007] Preferably, the first hybrid attention module and the second hybrid attention module are attention modules with the same structure or different structures, selected from CBAM, ECANet or coordinate attention module.
[0008] Preferably, the feature fusion in step S5 specifically involves: The high-level spatial features and the high-level spatiotemporal motion features are input into an adaptive feature fusion module. This module dynamically adjusts the contribution of the two features through learnable weight parameters, and then performs channel splicing or weighted summation.
[0009] Preferably, during the training phase, a phased training strategy is adopted for the time flow branches: First, freeze the temporal attention modeling unit and then train the feature extraction unit; Then unfreeze and fine-tune each time flow branch.
[0010] A vehicle collision detection system based on an improved two-stream convolutional network includes: The data acquisition module is configured to acquire raw video streams containing the target vehicle in real time via an onboard camera; The preprocessing module is configured to preprocess the original video stream to extract spatial frame sequences and optical flow frame sequences, respectively. Improved dual-stream convolutional network modules, including: The spatial flow branch, with MobileNetV3 as its backbone and embedded with the first hybrid attention module; The temporal branch includes a cascaded MobileNetV3 feature extraction sub-branch and a DW-ConvLSTM temporal attention modeling sub-branch; The feature fusion module is configured to fuse features extracted from two branches to generate a fused feature vector. The classification and output module is configured to input the fused feature vector into the classifier and output the collision detection result.
[0011] Preferably, the system is deployed in a hardware-software co-acceleration manner, wherein the MobileNetV3 backbone network and the DW-ConvLSTM layer run on an embedded AI chip that supports fixed-point arithmetic and hardware acceleration instruction sets.
[0012] It has the following beneficial effects: This solution achieves significant breakthroughs and collaborative optimizations across multiple dimensions in the field of vehicle collision detection, forming a comprehensive advantage of high precision, low latency, easy deployment, and strong reliability. By constructing a triple-enhancement architecture with a lightweight network as the backbone, integrating a hybrid attention mechanism and an innovative temporal modeling module, the system achieves a significant improvement in detection accuracy while maintaining a highly streamlined model and low power consumption. The end-to-end processing latency is extremely low, reserving ample warning and response time for active safety systems. Its dual-stream parallel processing and adaptive feature fusion design effectively balance computational efficiency and feature integrity, enabling the system to maintain a low false alarm rate in complex driving scenarios and clearly distinguish between emergency braking and real collision risks. Through modular design, model quantization, and phased training strategies, the solution ensures a smooth transition from algorithm development to vehicle deployment, fully meeting automotive-grade requirements. This technology not only significantly improves the reliability of collision prevention and reduces the overall cost of the system, but also provides a high-performance, stable, reliable, and rapidly mass-producible complete solution for intelligent vehicle active safety systems with its excellent real-time performance, outstanding adaptability, and good scalability, possessing significant industrial value and social significance. Detailed Implementation
[0013] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0014] In a first embodiment, the present invention provides a technical solution: a vehicle collision detection method based on an improved dual-stream convolutional network, comprising the following steps: Step 1: Dual-stream data preparation; Specific operation: Acquire a 1280×720@30fps video stream from the camera, extract one frame every two frames as spatial stream input, and simultaneously calculate the dense optical flow of 5 consecutive frames to generate a 10-frame optical flow sequence as temporal stream input; Spatial image sequences and optical flow sequences are generated in parallel to provide input data for the two-stream network; After the data preparation is completed, the two parallel data streams enter their respective feature extraction channels and begin the deep feature learning process. Step 2: Spatial feature extraction; Specific operation: Use the lightweight MobileNetV3-Small as the backbone network, embed the CBAM attention module in the key layer, and extract 576-dimensional spatial feature vectors from the image after standardization in step 1. By combining a lightweight CNN with a spatial attention mechanism, vehicle appearance, posture and environmental context features are efficiently extracted. While the spatial features are successfully extracted, the temporal data also completes the initial feature transformation, preparing for the next step of deep spatiotemporal fusion. Step 3: Spatiotemporal motion feature extraction; Specific operations: Level 1: Extract local motion features of optical flow for each frame using MobileNetV3+ coordinate attention; Level 2: Processes 10 consecutive frames of features through the innovative DW-ConvLSTM module, and models long-term motion dependencies using depthwise separable convolution and temporal attention; The two-level architecture first extracts local motion features, and then models long-term temporal patterns using a lightweight LSTM to form a complete understanding of motion. At this point, the system has obtained the spatial features representing the "static appearance" and the temporal features representing the "dynamic changes" respectively, and the two features will enter the fusion stage; Step 4: Adaptive feature fusion; Specific operation: Automatically learn the contribution weights of spatial flow and temporal flow (α=0.42, β=0.58), use weighted fusion instead of simple concatenation, and then classify through a fully connected network; The importance of dynamically balancing spatial appearance information and temporal motion information is emphasized to achieve more reasonable feature fusion. The fused feature vectors, which combine the essence of the two streams of information, will then be sent to the decision layer for final collision probability assessment. Step 5: Collision Decision and Response; Specific operation: Output the collision probability P_collision, set two response thresholds (0.7 and 0.9), and trigger different levels of early warning and active safety measures according to the probability value; The tiered response mechanism ensures timely early warning while reducing false alarms, and the delay is controlled within 30ms to meet real-time requirements. The decision result not only completes a single detection task, but also serves as system status feedback, influencing the processing strategy of subsequent frames and forming a continuously optimized detection loop.
[0015] Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art and related fields based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described and explained in the present invention, unless otherwise specified or limited, shall be implemented according to conventional means in the art.
Claims
1. A vehicle collision detection method based on an improved two-stream convolutional network, characterized in that, Includes the following steps: Step S1: Acquire raw video streams containing the target vehicle in real time using the vehicle-mounted camera; Step S2: Preprocess the original video stream to extract the spatial frame sequence and optical flow frame sequence respectively; Step S3: Input the spatial frame sequence into the spatial flow branch of the improved two-stream convolutional network, and input the optical flow frame sequence into the temporal flow branch of the improved two-stream convolutional network; The spatial flow branch uses the lightweight network MobileNetV3 as the feature extraction backbone and embeds a first hybrid attention module in at least one of its bottleneck layers. The temporal branch includes a feature extraction unit and a temporal attention modeling unit connected in sequence; the feature extraction unit uses the lightweight network MobileNetV3 as its backbone and embeds a second hybrid attention module in at least one of its bottleneck layers. The temporal attention modeling unit is a convolutional long short-term memory network module based on depthwise separable convolution, which is used to perform temporal attention weighting and long-term dependency modeling on the feature sequence output by the feature extraction unit. Step S4: Extract high-level spatial features and high-level spatiotemporal motion features through the spatial flow branch and the temporal flow branch, respectively; Step S5: Perform feature fusion on the extracted high-level spatial features and high-level spatiotemporal motion features to generate a fused feature vector; Step S6: Input the fused feature vector into the classifier and output the probability of the target vehicle colliding or the collision detection result.
2. The vehicle collision detection method based on an improved two-stream convolutional network according to claim 1, characterized in that, The temporal attention modeling unit specifically performs the following operations: Receive a feature map sequence of length T from the feature extraction unit; The feature map sequence is input into the DW-ConvLSTM layer, where the core computation uses depthwise separable convolution instead of standard convolution to reduce the number of parameters. Extract the hidden state of the DW-ConvLSTM layer at the last time step, or perform adaptive pooling on the hidden states of all time steps to obtain the high-level spatiotemporal motion features weighted by temporal attention.
3. The vehicle collision detection method based on an improved two-stream convolutional network according to claim 1, characterized in that, Following the DW-ConvLSTM layer, a lightweight spatiotemporal attention submodule is connected to further enhance the hidden state in the channel and spatial dimensions.
4. The vehicle collision detection method based on an improved two-stream convolutional network according to claim 1, characterized in that, The first hybrid attention module and the second hybrid attention module are attention modules with the same structure or different structures, selected from CBAM, ECANet or coordinate attention module.
5. The vehicle collision detection method based on an improved two-stream convolutional network according to claim 1, characterized in that, The feature fusion in step S5 specifically involves: The high-level spatial features and the high-level spatiotemporal motion features are input into an adaptive feature fusion module. This module dynamically adjusts the contribution of the two features through learnable weight parameters, and then performs channel splicing or weighted summation.
6. The vehicle collision detection method based on an improved two-stream convolutional network according to claim 1, characterized in that, During the training phase, a phased training strategy is adopted for the time flow branches: First, freeze the temporal attention modeling unit and then train the feature extraction unit; Then unfreeze and fine-tune each time stream branch.
7. A vehicle collision detection system based on an improved two-stream convolutional network, characterized in that, include: The data acquisition module is configured to acquire raw video streams containing the target vehicle in real time via an onboard camera; The preprocessing module is configured to preprocess the original video stream to extract spatial frame sequences and optical flow frame sequences, respectively. Improved dual-stream convolutional network modules, including: The spatial flow branch, with MobileNetV3 as its backbone and embedded with the first hybrid attention module; The temporal branch includes a cascaded MobileNetV3 feature extraction sub-branch and a DW-ConvLSTM temporal attention modeling sub-branch; The feature fusion module is configured to fuse features extracted from two branches to generate a fused feature vector. The classification and output module is configured to input the fused feature vector into the classifier and output the collision detection result.
8. A vehicle collision detection system based on an improved two-stream convolutional network according to claim 7, characterized in that, The system is deployed in a hardware-software co-acceleration manner, wherein the MobileNetV3 backbone network and the DW-ConvLSTM layer run on an embedded AI chip that supports fixed-point arithmetic and hardware acceleration instruction sets.